SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2104.03866 · CVPR · 2021

SMD-Nets: Stereo Mixture Density Networks

Andreas Geiger, Yiyi Liao, Fabio Tosi, Carolin Schmitt

arXiv · PDF · Open in the Atlas

Code that ran

15 of 28 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (3 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 10 have no contract verdict; 5 failures of our instrument; 13 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2104.03866.json · badge: /api/ran/2104.03866.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
Mirmix/ddlmvs community
Papers with Code archive
10 ran · 4 no recorded run
fabiotosi92/SMD-Nets marked official in Papers with Code
Papers with Code archive
5 ran · 9 no recorded run
FunctionStatusWhere it lives
ConvBnReLU3D Ran
ran on a synthesized input
Mirmix/ddlmvs/models/net.py
code served (permissive licence) · get_code("76a27023b936b897")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 5, dtype family float, shape (2, 4, 8, 8, 8) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import torch.nn.functional as F

class ConvBnReLU3D(nn.Module):
    """Implements of 3d convolution + batch normalization + ReLU."""

    def __init__(self, in_channels: int, out_channels: int, kernel_size: int=3, stride: int=1, pad: int=1, dilation: int=1) -> None:
        """initialization method for convolution3D + batch normalization + relu module
        Args:
            in_channels: input channel number of convolution layer
            out_channels: output channel number of convolution layer
            kernel_size: kernel size of convolution layer
            stride: stride of convolution layer
            pad: pad of convolution layer
            dilation: dilation of convolution layer
        """
        super(ConvBnReLU3D, self).__init__()
        self.conv = nn.Conv3d(in_channels, out_channels, kernel_size, stride=stride, padding=pad, dilation=dilation, bias=False)
        self.bn = nn.BatchNorm3d(out_channels)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """forward method"""
        return F.relu(self.bn(self.conv(x)), inplace=True)


INIT_ARGS = {
    "in_channels": 3,
    "out_channels": 4,
    "kernel_size": 3,
    "stride": 1,
    "pad": 1,
    "dilation": 1,
}

ARGS = {
    "x": torch.randn(2, 3, 8, 8, 8),
}

if __name__ == "__main__":
    module = ConvBnReLU3D(**INIT_ARGS)
    output = module(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Expected shape: (2, 4, 8, 8, 8)")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 5, dtype family float, shape (2, 4, 8, 8, 8) — shape (2, 4, 8, 8, 8) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
DepthInitialization Ran
ran on a synthesized input
Mirmix/ddlmvs/models/net.py
code served (permissive licence) · get_code("92bc6ab66d10346b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 4, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

class DepthInitialization(nn.Module):
    """Initialization Stage Class"""

    def __init__(self, patchmatch_num_sample: int=1) -> None:
        """Initialize method

        Args:
            patchmatch_num_sample: number of samples used in patchmatch process
        """
        super(DepthInitialization, self).__init__()
        self.patchmatch_num_sample = patchmatch_num_sample

    def forward(self, random_initialization: bool, min_depth: torch.Tensor, max_depth: torch.Tensor, height: int, width: int, depth_interval_scale: float, device: torch.device, depth: torch.Tensor=None) -> torch.Tensor:
        """Forward function for depth initialization

        Args:
            random_initialization: whether to use random initialization
            min_depth: minimum virtual depth, (B, )
            max_depth: maximum virtual depth, (B, )
            height: height of depth map
            width: width of depth map
            depth_interval_scale: depth interval scale
            device: device on which to place tensor
            depth: current depth (B, 1, H, W)

        Returns:
            depth_sample: initialized sample depth map by randomization or local perturbation (B, Ndepth, H, W)
        """
        batch_size = min_depth.size()[0]
        if random_initialization:
            inverse_min_depth = 1.0 / min_depth
            inverse_max_depth = 1.0 / max_depth
            patchmatch_num_sample = 48
            depth_sample = torch.rand(size=(batch_size, patchmatch_num_sample, height, width), device=device) + torch.arange(start=0, end=patchmatch_num_sample, step=1, device=device).view(1, patchmatch_num_sample, 1, 1)
            depth_sample = inverse_max_depth.view(batch_size, 1, 1, 1) + depth_sample / patchmatch_num_sample * (inverse_min_depth.view(batch_size, 1, 1, 1) - inverse_max_depth.view(batch_size, 1, 1, 1))
            depth_sample = 1.0 / depth_sample
            return depth_sample
        elif self.patchmatch_num_sample == 1:
            return depth.detach()
        else:
            inverse_min_depth = 1.0 / min_depth
            inverse_max_depth = 1.0 / max_depth
            depth_sample = torch.arange(-self.patchmatch_num_sample // 2, self.patchmatch_num_sample // 2, 1, device=device).view(1, self.patchmatch_num_sample, 1, 1).repeat(batch_size, 1, height, width).float()
            inverse_depth_interval = (inverse_min_depth - inverse_max_depth) * depth_interval_scale
            inverse_depth_interval = inverse_depth_interval.view(batch_size, 1, 1, 1)
            depth_sample = 1.0 / depth.detach() + inverse_depth_interval * depth_sample
            depth_clamped = []
            del depth
            for k in range(batch_size):
                depth_clamped.append(torch.clamp(depth_sample[k], min=inverse_max_depth[k], max=inverse_min_depth[k]).unsqueeze(0))
            depth_sample = 1.0 / torch.cat(depth_clamped, dim=0)
            del depth_clamped
            return depth_sample


INIT_ARGS = {
    'patchmatch_num_sample': 3
}

ARGS = {
    'random_initialization': True,
    'min_depth': torch.tensor([0.5, 0.6]),
    'max_depth': torch.tensor([5.0, 6.0]),
    'height': 4,
    'width': 3,
    'depth_interval_scale': 0.1,
    'device': torch.device('cpu'),
    'depth': torch.ones(2, 1, 4, 3)
}

model = DepthInitialization(**INIT_ARGS)
output = model(**ARGS)
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
FeatureNet Ran
ran on a synthesized input
Mirmix/ddlmvs/models/net.py
code served (permissive licence) · get_code("ee2cc279bdf6169f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Any, Dict, List

class ConvBnReLU(nn.Module):
    """Implements 2d Convolution + batch normalization + ReLU"""

    def __init__(self, in_channels: int, out_channels: int, kernel_size: int=3, stride: int=1, pad: int=1, dilation: int=1) -> None:
        """initialization method for convolution2D + batch normalization + relu module
        Args:
            in_channels: input channel number of convolution layer
            out_channels: output channel number of convolution layer
            kernel_size: kernel size of convolution layer
            stride: stride of convolution layer
            pad: pad of convolution layer
            dilation: dilation of convolution layer
        """
        super(ConvBnReLU, self).__init__()
        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride=stride, padding=pad, dilation=dilation, bias=False)
        self.bn = nn.BatchNorm2d(out_channels)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """forward method"""
        return F.relu(self.bn(self.conv(x)), inplace=True)

class FeatureNet(nn.Module):
    """Feature Extraction Network: to extract features of original images from each view"""

    def __init__(self):
        """Initialize different layers in the network"""
        super(FeatureNet, self).__init__()
        self.conv0 = ConvBnReLU(3, 8, 3, 1, 1)
        self.conv1 = ConvBnReLU(8, 8, 3, 1, 1)
        self.conv2 = ConvBnReLU(8, 16, 5, 2, 2)
        self.conv3 = ConvBnReLU(16, 16, 3, 1, 1)
        self.conv4 = ConvBnReLU(16, 16, 3, 1, 1)
        self.conv5 = ConvBnReLU(16, 32, 5, 2, 2)
        self.conv6 = ConvBnReLU(32, 32, 3, 1, 1)
        self.conv7 = ConvBnReLU(32, 32, 3, 1, 1)
        self.conv8 = ConvBnReLU(32, 64, 5, 2, 2)
        self.conv9 = ConvBnReLU(64, 64, 3, 1, 1)
        self.conv10 = ConvBnReLU(64, 64, 3, 1, 1)
        self.output1 = nn.Conv2d(64, 64, 1, bias=False)
        self.inner1 = nn.Conv2d(32, 64, 1, bias=True)
        self.inner2 = nn.Conv2d(16, 64, 1, bias=True)
        self.output2 = nn.Conv2d(64, 32, 1, bias=False)
        self.output3 = nn.Conv2d(64, 16, 1, bias=False)

    def forward(self, x: torch.Tensor) -> Dict[str, torch.Tensor]:
        """Forward method

        Args:
            x: images from a single view, in the shape of [B, C, H, W]. Generally, C=3

        Returns:
            output_feature: a python dictionary contains extracted features from stage_1 to stage_3
                keys are "stage_1", "stage_2", and "stage_3"
        """
        output_feature = {}
        conv1 = self.conv1(self.conv0(x))
        conv4 = self.conv4(self.conv3(self.conv2(conv1)))
        conv7 = self.conv7(self.conv6(self.conv5(conv4)))
        conv10 = self.conv10(self.conv9(self.conv8(conv7)))
        output_feature['stage_3'] = self.output1(conv10)
        intra_feat = F.interpolate(conv10, scale_factor=2, mode='bilinear') + self.inner1(conv7)
        del conv7, conv10
        output_feature['stage_2'] = self.output2(intra_feat)
        intra_feat = F.interpolate(intra_feat, scale_factor=2, mode='bilinear') + self.inner2(conv4)
        del conv4
        output_feature['stage_1'] = self.output3(intra_feat)
        del intra_feat
        return output_feature


INIT_ARGS = {}

ARGS = {
    'x': torch.randn(2, 3, 32, 32)
}

if __name__ == '__main__':
    model = FeatureNet(**INIT_ARGS)
    model.eval()
    output = model(**ARGS)
    print("Output keys:", output.keys())
    for key, val in output.items():
        print(f"{key}: shape {val.shape}")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
FeatureWeightNet Ran
ran on a synthesized input
Mirmix/ddlmvs/models/net.py
code served (permissive licence) · get_code("fde70f1c0e22a3e9")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 4, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

# INIT_ARGS for FeatureWeightNet.__init__
INIT_ARGS = {
    'num_feature': 32,  # number of features (must be divisible by G)
    'neighbors': 9,     # number of neighbors to be sampled
    'G': 8              # number of groups
}

# Create input tensors for forward method
batch = 2
feature_channel = 32  # must match num_feature
height = 8
width = 6

# ref_feature: [B, C, H, W]
ref_feature = torch.randn(batch, feature_channel, height, width)

# grid: position of sampling points [B, neighbors*H, W, 2]
# Grid values should be in [-1, 1] range for grid_sample
grid = torch.rand(batch, INIT_ARGS['neighbors'] * height, width, 2) * 2 - 1

# ARGS for FeatureWeightNet.forward
ARGS = {
    'ref_feature': ref_feature,
    'grid': grid
}
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
PixelwiseNet Ran
ran on a synthesized input
Mirmix/ddlmvs/models/net.py
code served (permissive licence) · get_code("de28fde5fb92de44")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 4, dtype family float, shape (2, 1, 5, 6) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import torch.nn.functional as F

class ConvBnReLU3D(nn.Module):
    """Implements of 3d convolution + batch normalization + ReLU."""

    def __init__(self, in_channels: int, out_channels: int, kernel_size: int=3, stride: int=1, pad: int=1, dilation: int=1) -> None:
        """initialization method for convolution3D + batch normalization + relu module
        Args:
            in_channels: input channel number of convolution layer
            out_channels: output channel number of convolution layer
            kernel_size: kernel size of convolution layer
            stride: stride of convolution layer
            pad: pad of convolution layer
            dilation: dilation of convolution layer
        """
        super(ConvBnReLU3D, self).__init__()
        self.conv = nn.Conv3d(in_channels, out_channels, kernel_size, stride=stride, padding=pad, dilation=dilation, bias=False)
        self.bn = nn.BatchNorm3d(out_channels)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """forward method"""
        return F.relu(self.bn(self.conv(x)), inplace=True)

class PixelwiseNet(nn.Module):
    """Pixelwise Net: A simple pixel-wise view weight network, composed of 1x1x1 convolution layers
    and sigmoid nonlinearities, takes the initial set of similarities to output a number between 0 and 1 per
    pixel as estimated pixel-wise view weight.

    1. The Pixelwise Net is used in adaptive evaluation step
    2. The similarity is calculated by ref_feature and other source_features warped by differentiable_warping
    3. The learned pixel-wise view weight is estimated in the first iteration of Patchmatch and kept fixed in the
    matching cost computation.
    """

    def __init__(self, G: int) -> None:
        """Initialize method

        Args:
            G: the feature channels of input will be divided evenly into G groups
        """
        super(PixelwiseNet, self).__init__()
        self.conv0 = ConvBnReLU3D(in_channels=G, out_channels=16, kernel_size=1, stride=1, pad=0)
        self.conv1 = ConvBnReLU3D(in_channels=16, out_channels=8, kernel_size=1, stride=1, pad=0)
        self.conv2 = nn.Conv3d(in_channels=8, out_channels=1, kernel_size=1, stride=1, padding=0)
        self.output = nn.Sigmoid()

    def forward(self, x1: torch.Tensor) -> torch.Tensor:
        """Forward method for PixelwiseNet

        Args:
            x1: pixel-wise view weight, [B, G, Ndepth, H, W], where G is the number of groups
        """
        x1 = self.conv2(self.conv1(self.conv0(x1))).squeeze(1)
        output = self.output(x1)
        del x1
        output = torch.max(output, dim=1)[0]
        return output.unsqueeze(1)


INIT_ARGS = {
    'G': 4
}

ARGS = {
    'x1': torch.randn(2, 4, 3, 5, 6)
}

if __name__ == '__main__':
    model = PixelwiseNet(**INIT_ARGS)
    model.eval()
    output = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Output: {output}")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float, shape (2, 1, 5, 6) — shape (2, 1, 5, 6) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
Propagation Ran
ran on a synthesized input
Mirmix/ddlmvs/models/net.py
code served (permissive licence) · get_code("9dc98895fe55dc6b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 4, dtype family float, shape (2, 24, 4, 3) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import torch.nn.functional as F

class Propagation(nn.Module):
    """ Propagation module implementation"""

    def __init__(self, neighbors: int=16) -> None:
        """Initialize method

        Args:
            neighbors: number of neighbors to be sampled in propagation
        """
        super(Propagation, self).__init__()
        self.neighbors = neighbors

    def forward(self, batch: int, height: int, width: int, depth_sample: torch.Tensor, grid: torch.Tensor, depth_min: torch.Tensor, depth_max: torch.Tensor, depth_interval_scale: float) -> torch.Tensor:
        """Forward method of adaptive propagation

        Args:
            batch: batch size,
            height: depth map height,
            width: depth map width,
            depth_sample: sample depth map, in shape of [batch, num_depth, height, width],
            grid: 2D grid for bilinear gridding, in shape of [batch, neighbors*H, W, 2]
            depth_min: minimum virtual depth, in shape of [batch, ]
            depth_max: maximum virtual depth, in shape of [batch, ]
            depth_interval_scale: depth virtual interval scale,

        Returns:
            propagate depth: sorted propagate depth map [batch, num_depth+num_neighbors, height, width]
        """
        num_depth = depth_sample.size()[1]
        propagate_depth = depth_sample.new_empty(batch, num_depth + self.neighbors, height, width)
        propagate_depth[:, 0:num_depth, :, :] = depth_sample
        propagate_depth_sample = F.grid_sample(depth_sample[:, num_depth // 2, :, :].unsqueeze(1), grid, mode='bilinear', padding_mode='border')
        del grid
        propagate_depth_sample = propagate_depth_sample.view(batch, self.neighbors, height, width)
        propagate_depth[:, num_depth:, :, :] = propagate_depth_sample
        del propagate_depth_sample
        propagate_depth, _ = torch.sort(propagate_depth, dim=1)
        return propagate_depth


INIT_ARGS = {
    'neighbors': 16
}

batch_size = 2
height = 4
width = 3
num_depth = 8
neighbors = 16

ARGS = {
    'batch': batch_size,
    'height': height,
    'width': width,
    'depth_sample': torch.randn(batch_size, num_depth, height, width),
    'grid': torch.randn(batch_size, neighbors * height, width, 2),
    'depth_min': torch.randn(batch_size),
    'depth_max': torch.randn(batch_size),
    'depth_interval_scale': 1.0
}

if __name__ == '__main__':
    model = Propagation(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Expected shape: ({batch_size}, {num_depth + neighbors}, {height}, {width})")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float, shape (2, 24, 4, 3) — shape (2, 24, 4, 3) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
Regressor Ran
ran on a synthesized input · constructed an object; no result was computed
fabiotosi92/SMD-Nets/lib/model/SMDHead.py
code served (permissive licence) · get_code("c2896a15e4b9aa7d")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, Regressor: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Analyze the Regressor class:
# - __init__ takes: filter_channels, no_sine, no_residual
# - filter_channels is used in: len(filter_channels), indexing filter_channels[l] and filter_channels[l+1]
# - no_sine is used in: conditional for activation function
# - no_residual is used in: conditionals in __init__ and forward
# - forward takes feature tensor and applies convolutions with optional residual connections
# - feature tensor shape: (batch, channels, length) for Conv1d
#   - channels must match filter_channels[0] initially
#   - length can be any positive value

# Construct filter_channels: list of channel dimensions
# Start with some channel size, then transition through layers
# Must have at least 2 elements (start and end)
filter_channels = [4, 8, 3]  # 4 input channels, 8 hidden, 3 output

# Create a small test feature tensor
# Shape: (batch_size=2, in_channels=4, length=5)
feature = torch.randn(2, 4, 5)

# Test with both no_sine and no_residual combinations
ARGS = {
    'filter_channels': filter_channels,
    'no_sine': False,
    'no_residual': False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Regressor — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a Regressor output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
ResNetMultiImageInput Ran
ran on a synthesized input · constructed an object; no result was computed
fabiotosi92/SMD-Nets/lib/model/SMDHead.py
code served (permissive licence) · get_code("0e653694acc4e981")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, ResNetMultiImageInput: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import torchvision.models as models

torch.manual_seed(0)

# The ResNetMultiImageInput.__init__ calls:
# - super().__init__(block, layers) which expects ResNet block class and layer counts
# - self._make_layer() which is inherited from ResNet and uses the block class
# - Operations on self.modules() for initialization

# We need to provide:
# - block: a ResNet block class (BasicBlock or Bottleneck)
# - layers: list/tuple of 4 integers specifying layer depths
# - num_classes: int (default 1000)
# - num_input_images: int (default 1)
# - channels: int (default 1)

# Use BasicBlock from torchvision as the block
from torchvision.models.resnet import BasicBlock

ARGS = {
    "block": BasicBlock,
    "layers": [2, 2, 2, 2],
    "num_classes": 1000,
    "num_input_images": 1,
    "channels": 1,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, torchvision — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ResNetMultiImageInput — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ResNetMultiImageInput output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
SimilarityNet Ran
ran on a synthesized input
Mirmix/ddlmvs/models/net.py
code served (permissive licence) · get_code("52e4096f56488419")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 4, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import torch.nn.functional as F

class ConvBnReLU3D(nn.Module):
    """Implements of 3d convolution + batch normalization + ReLU."""

    def __init__(self, in_channels: int, out_channels: int, kernel_size: int=3, stride: int=1, pad: int=1, dilation: int=1) -> None:
        """initialization method for convolution3D + batch normalization + relu module
        Args:
            in_channels: input channel number of convolution layer
            out_channels: output channel number of convolution layer
            kernel_size: kernel size of convolution layer
            stride: stride of convolution layer
            pad: pad of convolution layer
            dilation: dilation of convolution layer
        """
        super(ConvBnReLU3D, self).__init__()
        self.conv = nn.Conv3d(in_channels, out_channels, kernel_size, stride=stride, padding=pad, dilation=dilation, bias=False)
        self.bn = nn.BatchNorm3d(out_channels)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """forward method"""
        return F.relu(self.bn(self.conv(x)), inplace=True)

class SimilarityNet(nn.Module):
    """Similarity Net, used in Evaluation module (adaptive evaluation step)
    1. Do 1x1x1 convolution on aggregated cost [B, G, Ndepth, H, W] among all the source views,
        where G is the number of groups
    2. Perform adaptive spatial cost aggregation to get final cost (scores)
    """

    def __init__(self, G: int, neighbors: int=9) -> None:
        """Initialize method

        Args:
            G: the feature channels of input will be divided evenly into G groups
            neighbors: number of neighbors to be sampled
        """
        super(SimilarityNet, self).__init__()
        self.neighbors = neighbors
        self.conv0 = ConvBnReLU3D(in_channels=G, out_channels=16, kernel_size=1, stride=1, pad=0)
        self.conv1 = ConvBnReLU3D(in_channels=16, out_channels=8, kernel_size=1, stride=1, pad=0)
        self.similarity = nn.Conv3d(in_channels=8, out_channels=1, kernel_size=1, stride=1, padding=0)

    def forward(self, x1: torch.Tensor, grid: torch.Tensor, weight: torch.Tensor) -> torch.Tensor:
        """Forward method for SimilarityNet

        Args:
            x1: [B, G, Ndepth, H, W], where G is the number of groups, aggregated cost among all the source views with
                pixel-wise view weight
            grid: position of sampling points in adaptive spatial cost aggregation, (B, evaluate_neighbors*H, W, 2)
            weight: weight of sampling points in adaptive spatial cost aggregation, combination of
                feature weight and depth weight, [B,Ndepth,1,H,W]

        Returns:
            final cost: in the shape of [B,Ndepth,H,W]
        """
        batch, G, num_depth, height, width = x1.size()
        x1 = self.similarity(self.conv1(self.conv0(x1))).squeeze(1)
        x1 = F.grid_sample(x1, grid, mode='bilinear', padding_mode='border')
        x1 = x1.view(batch, num_depth, self.neighbors, height, width)
        return torch.sum(x1 * weight, dim=2)

INIT_ARGS = {
    'G': 8,
    'neighbors': 9
}

batch = 2
G = 8
num_depth = 4
height = 6
width = 5
neighbors = 9

x1 = torch.randn(batch, G, num_depth, height, width)
grid = torch.randn(batch, neighbors * height, width, 2)
weight = torch.randn(batch, num_depth, neighbors, height, width)

ARGS = {
    'x1': x1,
    'grid': grid,
    'weight': weight
}
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
Sine Ran
ran on a synthesized input · constructed an object; no result was computed
fabiotosi92/SMD-Nets/lib/model/SMDHead.py
code served (permissive licence) · get_code("a16de707e5586b16")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, Sine: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

ARGS = {
    "w0": 2.0
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Sine — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a Sine output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
depth_regression Ran
ran; our contract draft was wrong, not the code
Mirmix/ddlmvs/models/net.py
code served (permissive licence) · get_code("e8f9986089f1f048")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned Tensor, ndim 4, dtype family float, shape (2, 1, 3, 5) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Construct tiny tensors for depth regression
batch_size = 2
num_depth_values = 4
height = 3
width = 5

# p: probability volume [B, D, H, W]
# Should contain probabilities that sum to 1 along the D dimension
p = torch.rand(batch_size, num_depth_values, height, width)
p = p / p.sum(dim=1, keepdim=True)  # normalize to sum to 1 along depth dimension

# depth_values: discrete depth values [B, D]
depth_values = torch.linspace(0.5, 5.0, num_depth_values).unsqueeze(0).expand(batch_size, -1)

ARGS = {
    "p": p,
    "depth_values": depth_values,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float, shape (2, 1, 3, 5) — shape (2, 1, 3, 5) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
depth_weight Ran
ran; our contract draft was wrong, not the code
Mirmix/ddlmvs/models/net.py
code served (permissive licence) · get_code("25252e58e3b331de")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned Tensor, ndim 5, dtype family float, shape (2, 4, 3, 3, 5) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Define tensor dimensions
batch = 2
num_depth = 4
height = 3
width = 5

# depth_sample: (B, Ndepth, H, W)
depth_sample = torch.randn(batch, num_depth, height, width) + 1.0  # Add offset to avoid near-zero values

# depth_min: (B,)
depth_min = torch.tensor([0.5, 0.6])

# depth_max: (B,)
depth_max = torch.tensor([5.0, 6.0])

# grid: (B, evaluate_neighbors*H, W, 2)
# This is used with F.grid_sample, so it should contain normalized coordinates in [-1, 1]
evaluate_neighbors = 3
grid = torch.rand(batch, evaluate_neighbors * height, width, 2) * 2 - 1  # Normalized to [-1, 1]

# patchmatch_interval_scale: float
patchmatch_interval_scale = 0.5

# evaluate_neighbors: int
evaluate_neighbors = 3

ARGS = {
    "depth_sample": depth_sample,
    "depth_min": depth_min,
    "depth_max": depth_max,
    "grid": grid,
    "patchmatch_interval_scale": patchmatch_interval_scale,
    "evaluate_neighbors": evaluate_neighbors,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 5, dtype family float, shape (2, 4, 3, 3, 5) — shape (2, 4, 3, 3, 5) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
differentiable_warping Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
Mirmix/ddlmvs/models/net.py
code served (permissive licence) · get_code("00b8d4f5fece378f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned Tensor, ndim 5, dtype family float, shape (2, 4, 3, 8, 6) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn.functional as F

torch.manual_seed(0)

# Define dimensions
batch_size = 2
channels = 4
height = 8
width = 6
num_depth = 3

# src_fea: [B, C, H, W] source features
src_fea = torch.randn(batch_size, channels, height, width)

# src_proj: [B, 4, 4] source camera projection matrix
src_proj = torch.randn(batch_size, 4, 4)

# ref_proj: [B, 4, 4] reference camera projection matrix
ref_proj = torch.randn(batch_size, 4, 4)
# Ensure ref_proj is invertible by making it closer to identity
ref_proj = ref_proj + torch.eye(4).unsqueeze(0) * 2

# depth_samples: [B, Ndepth, H, W] virtual depth layers
depth_samples = torch.randn(batch_size, num_depth, height, width).abs() + 0.5

ARGS = {
    "src_fea": src_fea,
    "src_proj": src_proj,
    "ref_proj": ref_proj,
    "depth_samples": depth_samples,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 5, dtype family float, shape (2, 4, 3, 8, 6) — shape (2, 4, 3, 8, 6) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
local_contrast_normalization Ran
ran; our contract draft was wrong, not the code
fabiotosi92/SMD-Nets/lib/model/SMDHead.py
code served (permissive licence) · get_code("eb4a51cd37197a6e")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 4, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create a small image tensor
# Shape: (batch_size, height, width) - 2D image with batch dimension
image = torch.randn(2, 4, 3)

ARGS = {
    "image": image,
    "radius": 25,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 4, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
resnet_multiimage_input Ran
ran; our contract draft was wrong, not the code
fabiotosi92/SMD-Nets/lib/model/SMDHead.py
code served (permissive licence) · get_code("98f5a5eb6fec0a91")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ResNetMultiImageInput. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import torchvision.models as models

torch.manual_seed(0)

ARGS = {
    'num_layers': 18,
    'pretrained': False,
    'num_input_images': 1,
    'channels': 1,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, torchvision — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ResNetMultiImageInput — a ResNetMultiImageInput output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a ResNetMultiImageInput output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
DDLMVS Not yet run
unverified (harvested, no recorded run)
Mirmix/ddlmvs/models/net.py
code served (permissive licence) · get_code("89bd4abf9c72da4f")
How it ran

no invocation recorded.

DepthDecoder Not yet run
unverified (harvested, no recorded run)
fabiotosi92/SMD-Nets/lib/model/SMDHead.py
code served (permissive licence) · get_code("7ccf369a6cafb07a")
How it ran

no invocation recorded.

Evaluation Not yet run
unverified (harvested, no recorded run)
Mirmix/ddlmvs/models/net.py
code served (permissive licence) · get_code("c0aa140f64e3d229")
How it ran

no invocation recorded.

HSMNet Not yet run
unverified (harvested, no recorded run)
fabiotosi92/SMD-Nets/lib/model/SMDHead.py
code served (permissive licence) · get_code("338810fc5e2043a8")
How it ran

no invocation recorded.

PSMNet Not yet run
unverified (harvested, no recorded run)
fabiotosi92/SMD-Nets/lib/model/SMDHead.py
code served (permissive licence) · get_code("15bbcc83e8f87378")
How it ran

no invocation recorded.

PatchMatch Not yet run
unverified (harvested, no recorded run)
Mirmix/ddlmvs/models/net.py
code served (permissive licence) · get_code("9c7781a0f6f917c4")
How it ran

no invocation recorded.

Refinement Not yet run
unverified (harvested, no recorded run)
Mirmix/ddlmvs/models/net.py
code served (permissive licence) · get_code("117934fc23392abd")
How it ran

no invocation recorded.

ResnetEncoder Not yet run
unverified (harvested, no recorded run)
fabiotosi92/SMD-Nets/lib/model/SMDHead.py
code served (permissive licence) · get_code("4af63eb5010aaaf7")
How it ran

no invocation recorded.

SDFilter Not yet run
unverified (harvested, no recorded run)
fabiotosi92/SMD-Nets/lib/model/SMDHead.py
code served (permissive licence) · get_code("12cf0361c050441b")
How it ran

no invocation recorded.

SMDHead Not yet run
unverified (harvested, no recorded run)
fabiotosi92/SMD-Nets/lib/model/SMDHead.py
code served (permissive licence) · get_code("c8a3ac84df7ddea3")
How it ran

no invocation recorded.

gaussian_blur Not yet run
unverified (harvested, no recorded run)
fabiotosi92/SMD-Nets/lib/model/SMDHead.py
code served (permissive licence) · get_code("e2fb247c688205e5")
How it ran

no invocation recorded.

get_backbone Not yet run
unverified (harvested, no recorded run)
fabiotosi92/SMD-Nets/lib/model/SMDHead.py
code served (permissive licence) · get_code("23b44af6dc3ed7f3")
How it ran

no invocation recorded.

hourglass Not yet run
unverified (harvested, no recorded run)
fabiotosi92/SMD-Nets/lib/model/SMDHead.py
code served (permissive licence) · get_code("47c7f7aa6bec2802")
How it ran

no invocation recorded.

Also present in other papers' code

Of 28 samples lifted from this paper's linked repositories, 28 carry a function identity and 0 do not. Of the 28 distinct functions with an identity, 2 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 2 later) and 26 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
Sine
brick of this paper's assembly: helper
8 other papers
1 dated earlier, 7 later · 4 only as an assembly brick
2006.09661 (2020-06)
Implicit Neural Representations with Periodic Activation Functions
2305.19185 (2023-05)
Compression with Bayesian Implicit Neural Representations
2310.19629 (2023-10)
RayDF: Neural Ray-surface Distance Fields with Multi-view Consistency
and 5 more not named here
hourglass
brick of this paper's assembly: helper
1 other paper
1 later · 1 only as an assembly brick
2407.10831 (2024-07)
Temporal Event Stereo via Joint Learning with Stereoscopic Flow

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Despite stereo matching accuracy has greatly improved by deep learning in the last few years, recovering sharp boundaries and high-resolution outputs efficiently remains challenging. In this paper, we propose Stereo Mixture Density Networks (SMD-Nets), a simple yet effective learning framework compatible with a wide class of 2D and 3D architectures which ameliorates both issues. Specifically, we exploit bimodal mixture densities as output representation and show that this allows for sharp and precise disparity estimates near discontinuities while explicitly modeling the aleatoric uncertainty inherent in the observations. Moreover, we formulate disparity estimation as a continuous problem in the image domain, allowing our model to query disparities at arbitrary spatial precision. We carry out comprehensive experiments on a new high-resolution and highly realistic synthetic stereo dataset, consisting of stereo pairs at 8Mpx resolution, as well as on real-world stereo datasets. Our experiments demonstrate increased depth accuracy near object boundaries and prediction of ultra high-resolution disparity maps on standard GPUs. We demonstrate the flexibility of our technique by improving the performance of a variety of stereo backbones.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2104.03866")
get_code_for_paper("2104.03866")
have("2104.03866")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2104.03866.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2104.03866.svg)](https://syntology.ai/paper/2104.03866)

Connect an agent — have() is free.