SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2409.03757 · NeurIPS · 2024

Lexicon3D: Probing Visual Foundation Models for Complex 3D Scene Understanding

Zhipeng Bao, Yu-Xiong Wang, Martial Hebert, Liang-Yan Gui, Yunze Man, Shuhong Zheng

arXiv · PDF · Open in the Atlas

Code that ran

8 of 9 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 0 violated it and 6 have no contract verdict; 1 failure of our instrument; 1 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/2409.03757.json · badge: /api/ran/2409.03757.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
yunzeman/lexicon3d community (classifier label)
classifier basis not recorded on the link
8 ran · 1 no recorded run
FunctionStatusWhere it lives
adjust_intrinsic Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
yunzeman/lexicon3d/lexicon3d/fusion_util.py
code served (permissive licence) · get_code("c58cf40ab0e3d808")
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 Tensor, ndim 2, dtype family float, shape (3, 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 math

torch.manual_seed(0)

# The function adjust_intrinsic takes:
# - intrinsic: a tensor (camera intrinsic matrix, 3x3 is typical but function treats it as indexable [0,0], [1,1], [0,2], [1,2])
# - intrinsic_image_dim: tuple of (height, width) of original image
# - image_dim: tuple of (height, width) of target image

# Create a 3x3 intrinsic matrix (camera intrinsics)
intrinsic = torch.tensor([
    [500.0, 0.0, 320.0],
    [0.0, 500.0, 240.0],
    [0.0, 0.0, 1.0]
], dtype=torch.float32)

# Original image dimensions (height, width)
intrinsic_image_dim = (480, 640)

# Target image dimensions (height, width) - different from original
image_dim = (240, 320)

ARGS = {
    "intrinsic": intrinsic,
    "intrinsic_image_dim": intrinsic_image_dim,
    "image_dim": image_dim,
}
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 Tensor, ndim 2, dtype family float, shape (3, 3) — shape (3, 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: none by the harness; any seed is in the setup block
extract_lseg_img_feature Ran
ran on a synthesized input
yunzeman/lexicon3d/lexicon3d/fusion_util.py
code served (permissive licence) · get_code("6d3d570ec9c749ac")
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 Tensor, ndim 3, dtype family float, shape (8, 8, 4) 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 numpy as np
from PIL import Image
import io

torch.manual_seed(0)

# Create a minimal transform that converts numpy array to tensor
class MinimalTransform:
    def __call__(self, image):
        # Convert numpy array to tensor
        if isinstance(image, np.ndarray):
            # Handle different image formats
            if len(image.shape) == 2:  # grayscale
                image = np.stack([image] * 3, axis=-1)
            image = torch.from_numpy(image).float()
            # Normalize to [0, 1] if needed and permute to CHW
            if image.max() > 1:
                image = image / 255.0
            if len(image.shape) == 3 and image.shape[2] == 3:
                image = image.permute(2, 0, 1)
            return image
        return image

# Create a minimal evaluator with parallel_forward method
class MinimalEvaluator(nn.Module):
    def __init__(self):
        super().__init__()
        # Small dummy layer to satisfy nn.Module
        self.dummy = nn.Linear(4, 4)
    
    def parallel_forward(self, image, label=''):
        # image shape: (batch, channels, height, width)
        # Return format: list containing list with tensor
        batch_size = image.shape[0]
        # Return a feature tensor of shape (height, width, features)
        feat = torch.randn(8, 8, 4, dtype=torch.float32)
        return [[feat]]

# Create a minimal image file (RGB, 16x12 pixels)
img_array = np.random.randint(0, 256, (12, 16, 3), dtype=np.uint8)
img = Image.fromarray(img_array, mode='RGB')
img_bytes = io.BytesIO()
img.save(img_bytes, format='PNG')
img_bytes.seek(0)
img_path = img_bytes

# Create stand-in for img_dir - we'll use a temporary in-memory approach
# Since we can't do file I/O, create a mock that returns image data
class MockImagePath:
    def __init__(self, img_array):
        self.img_array = img_array
    
    def __str__(self):
        return "mock_image.png"

mock_img = MockImagePath(img_array)

# Override Image.open to return our test image
original_open = Image.open
def mock_open(path):
    if isinstance(path, MockImagePath):
        return Image.fromarray(path.img_array, mode='RGB')
    return original_open(path)

Image.open = mock_open

ARGS = {
    'img_dir': mock_img,
    'transform': MinimalTransform(),
    'evaluator': MinimalEvaluator(),
    'label': 'test_label'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, 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 3, dtype family float, shape (8, 8, 4) — shape (8, 8, 4) 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
make_intrinsic Ran
ran, honoured the contract we drafted
yunzeman/lexicon3d/lexicon3d/fusion_util.py
code served (permissive licence) · get_code("6385efe2f5a14a24")
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 ndarray, ndim 2, dtype family float, shape (4, 4) 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 numpy as np

# The function creates a 4x4 intrinsic camera matrix
# Parameters are focal lengths (fx, fy) and principal point (mx, my)
# All are scalar floats representing camera calibration values

ARGS = {
    "fx": 500.0,
    "fy": 500.0,
    "mx": 320.0,
    "my": 240.0,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 2, dtype family float, shape (4, 4) — shape (4, 4) 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
measure_inference_time Ran
ran on a synthesized input
yunzeman/lexicon3d/lexicon3d/evaluate_complexity.py
code served (permissive licence) · get_code("d75e167c843e048a")
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 float, ndim 0, dtype family float, shape () 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)

# Create mock models for each model_type
class DINOv2Model(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 8)
    
    def forward_features(self, x):
        return self.linear(x)

class LSEGModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 8)
    
    def parallel_forward(self, x, arg):
        return self.linear(x)

class SDModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 8)
    
    def forward(self, x, t=None, up_ft_index=None):
        return self.linear(x)

class CLIPModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 8)
    
    def __call__(self, x):
        return self.linear(x)

class VJEPAModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 8)
    
    def __call__(self, x):
        return self.linear(x)

class SVDModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 8)
    
    def forward(self, x, y, t=None, up_ft_index=None):
        return self.linear(x)

# Create a simple model for testing
model = DINOv2Model()

# Create input tensor: batch_size=2, features=4
input_tensor = torch.randn(2, 4)

ARGS = {
    'model': model,
    'model_type': 'dinov2',
    'input_tensor': input_tensor,
    'num_steps': 10
}
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 float, ndim 0, dtype family float, shape () — shape () 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
module_inference Ran
ran on a synthesized input
yunzeman/lexicon3d/lexicon3d/models.py
code served (permissive licence) · get_code("68cce1e838d6d71a")
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 Tensor, ndim 4, dtype family float, shape (2, 3, 4, 5) as a later run recorded it. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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)

class MockModule(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 5)
    
    def evaluate_random(self, image, label_set):
        # image shape: (batch, channels, height, width)
        # return a tensor of same shape
        batch_size = image.size(0)
        return torch.randn(batch_size, 3, 4, 5)

module = MockModule()
image = torch.randn(2, 3, 4, 5)
label_set = ["class_a", "class_b", "class_c"]
flip = True

ARGS = {
    "module": module,
    "image": image,
    "label_set": label_set,
    "flip": flip
}
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 Tensor, ndim 4, dtype family float, shape (2, 3, 4, 5) — shape (2, 3, 4, 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
pad_image Ran
ran on a synthesized input
yunzeman/lexicon3d/lexicon3d/models.py
code served (permissive licence) · get_code("28ad88e75cdc78c6")
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 Tensor, ndim 4, dtype family float, shape (2, 3, 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 numpy as np

torch.manual_seed(0)

# Create a small test image tensor: batch=2, channels=3, height=4, width=5
img = torch.randn(2, 3, 4, 5)

# Define mean and std as numpy arrays for normalization
mean = np.array([0.485, 0.456, 0.406])
std = np.array([0.229, 0.224, 0.225])

# Set crop_size larger than image dimensions to trigger padding
crop_size = 8

ARGS = {
    "img": img,
    "mean": mean,
    "std": std,
    "crop_size": crop_size
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
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, shape (2, 3, 8, 8) — shape (2, 3, 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: none by the harness; any seed is in the setup block
resize_image Ran
ran on a synthesized input
yunzeman/lexicon3d/lexicon3d/models.py
code served (permissive licence) · get_code("2a9ddb0649c62fe4")
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 Tensor, ndim 4, dtype family float, shape (2, 4, 6, 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.functional as F

torch.manual_seed(0)

# Create a small image tensor: batch_size=2, channels=4, height=3, width=5
img = torch.randn(2, 4, 3, 5)

# Target height and width
h = 6
w = 8

# up_kwargs is already defined in the function, but we can pass additional kwargs
# The function signature shows **up_kwargs, which means we pass keyword arguments
up_kwargs_args = {'mode': 'bilinear', 'align_corners': True}

ARGS = {
    'img': img,
    'h': h,
    'w': w,
    'mode': 'bilinear',
    'align_corners': True
}
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 Tensor, ndim 4, dtype family float, shape (2, 4, 6, 8) — shape (2, 4, 6, 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: none by the harness; any seed is in the setup block
run_model Ran
ran on a synthesized input
yunzeman/lexicon3d/lexicon3d/evaluate_complexity.py
code served (permissive licence) · get_code("0781f631b94b8147")
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 Tensor, ndim 3, dtype family float, shape (2, 4, 16) 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)

# Create mock models for each model_type
class DINOv2Mock(nn.Module):
    def forward_features(self, x):
        return torch.randn(2, 4, 16)

class LSEGMock(nn.Module):
    def parallel_forward(self, x, text):
        return torch.randn(2, 4, 8)

class SDMock(nn.Module):
    def forward(self, x, t, up_ft_index):
        return torch.randn(2, 4, 8)

class CLIPMock(nn.Module):
    def __call__(self, x):
        return torch.randn(2, 4, 8)

class VJEPAMock(nn.Module):
    def __call__(self, x):
        return torch.randn(2, 4, 8)

class SVDMock(nn.Module):
    def forward(self, x, y, t, up_ft_index):
        return torch.randn(2, 4, 8)

# Create a small input tensor (batch=2, channels=3, height=4, width=4)
input_tensor = torch.randn(2, 3, 4, 4)

# Create models for each type
dinov2_model = DINOv2Mock()
lseg_model = LSEGMock()
sd_model = SDMock()
clip_model = CLIPMock()
vjepa_model = VJEPAMock()
svd_model = SVDMock()

# Test with one model type (dinov2)
ARGS = {
    'model': dinov2_model,
    'model_type': 'dinov2',
    'input_tensor': input_tensor
}
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 Tensor, ndim 3, dtype family float, shape (2, 4, 16) — shape (2, 4, 16) 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
measure_memory Not yet run
unverified (harvested, no recorded run)
yunzeman/lexicon3d/lexicon3d/evaluate_complexity.py
code served (permissive licence) · get_code("3f9caaede747a476")
How it ran

no invocation recorded.

Also present in other papers' code

Of 9 samples lifted from this paper's linked repositories, 9 carry a function identity and 0 do not. Of the 9 distinct functions with an identity, 4 are also present in other papers whose code we hold (4 with at least one paper dated earlier, 0 the same month, 2 later) and 5 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
adjust_intrinsic 4 other papers
3 dated earlier, 1 later
2205.13490 (2022-05)
SemAffiNet: Semantic-Affine Transformation for Point Cloud Segmentation
2311.17707 (2023-11)
SAMPro3D: Locating SAM Prompts in 3D for Zero-Shot Instance Segmentation
2312.10671 (2023-12)
Open3DIS: Open-Vocabulary 3D Instance Segmentation with 2D Mask Guidance
and 1 more not named here
make_intrinsic 4 other papers
3 dated earlier, 1 later
2205.13490 (2022-05)
SemAffiNet: Semantic-Affine Transformation for Point Cloud Segmentation
2311.17707 (2023-11)
SAMPro3D: Locating SAM Prompts in 3D for Zero-Shot Instance Segmentation
2312.10671 (2023-12)
Open3DIS: Open-Vocabulary 3D Instance Segmentation with 2D Mask Guidance
and 1 more not named here
pad_image 1 other paper
1 dated earlier
2210.05714 (2022-10)
Visual Language Maps for Robot Navigation
resize_image 1 other paper
1 dated earlier
2210.05714 (2022-10)
Visual Language Maps for Robot Navigation

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

Complex 3D scene understanding has gained increasing attention, with scene encoding strategies built on top of visual foundation models playing a crucial role in this success. However, the optimal scene encoding strategies for various scenarios remain unclear, particularly compared to their image-based counterparts. To address this issue, we present the first comprehensive study that probes various visual encoding models for 3D scene understanding, identifying the strengths and limitations of each model across different scenarios. Our evaluation spans seven vision foundation encoders, including image, video, and 3D foundation models. We evaluate these models in four tasks: Vision-Language Scene Reasoning, Visual Grounding, Segmentation, and Registration, each focusing on different aspects of scene understanding. Our evaluation yields key intriguing findings: Unsupervised image foundation models demonstrate superior overall performance, video models excel in object-level tasks, diffusion models benefit geometric tasks, languagepretrained models show unexpected limitations in language-related tasks, and the mixture-of-vision-expert (MoVE) strategy leads to consistent performance improvement. These insights challenge some conventional understandings, provide novel perspectives on leveraging visual foundation models, and highlight the need for more flexible encoder selection in future vision-language and scene understanding tasks.

For agents

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

get_harvested_code_for_paper("2409.03757")
get_code_for_paper("2409.03757")
have("2409.03757")

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

curl https://syntology.ai/api/ran/2409.03757.json

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

[![Syntology run record](https://syntology.ai/api/ran/2409.03757.svg)](https://syntology.ai/paper/2409.03757)

Connect an agent — have() is free.