Dataset Card for TransPhy3D

TransPhy3D (FiftyOne) is a grouped FiftyOne dataset built from a random subset of the TransPhy3D test split. Each group contains one RGB video sequence plus a merged 3D reconstruction of the same scene.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from huggingface_hub import snapshot_download


# Download the dataset snapshot to the current working directory

snapshot_download(
    repo_id="Voxel51/TransPhy3D", 
    local_dir=".", 
    repo_type="dataset"
    )

# Load dataset from current directory using FiftyOne's native format
dataset = fo.Dataset.from_dir(
    dataset_dir=".",  # Current directory contains the dataset files
    dataset_type=fo.types.FiftyOneDataset,  # Specify FiftyOne dataset format
    name="TransPhy3D"  # Assign a name to the dataset for identification
)

# Launch the App
session = fo.launch_app(dataset)

Dataset Details

Dataset Description

This is a FiftyOne dataset with 51 samples.

The source dataset provides synthetic transparent/reflective object scenes rendered in Blender/Cycles with RGB video frames, 16-bit depth maps, surface normal maps, and per-frame camera calibration.

• Source dataset: Daniellesry/TransPhy3D
• Source split used: test
• License: Apache 2.0 (source dataset license)

Dataset Sources

• Repository: https://huggingface.co/datasets/Daniellesry/TransPhy3D
• Paper: Diffusion Knows Transparency (arXiv:2512.23705)
• Project page: https://daniellli.github.io/projects/DKT/
• GitHub: https://github.com/Daniellli/DKT

---

Sampling From the Source Test Set

The full TransPhy3D test split contains 131 WebDataset tar sequences.

This FiftyOne dataset uses a random subset of 51 sequences (~39% of the test split):

Step	Count	Notes
Initial random sample	33	~25% of 131, random.seed(42)
Additional sequences	18	disjoint from the first 33, same seed
Total in this dataset	51	listed in data/selected_tars.txt

Scene type breakdown in this sample:

• 39 materials sequences (0826_*)
• 12 table_with_robot sequences (0827_table_with_robot_*)

Each selected sequence contains 121 frames.

---

FiftyOne Dataset Structure

Dataset name: TransPhy3D

Media type: group

Default group slice: video

Groups and slices

There are 51 groups. Each group has two linked slices:

Slice	Media type	filepath	Description
video	video	rgb.mp4	RGB video assembled from source frames
reconstruction	3d	scene.fo3d	Merged RGB-colored point cloud for the whole scene

Switch slices in the FiftyOne App to view the video annotations or the 3D reconstruction for the same sequence.

Sample-level fields

Present on both slices unless noted:

Field	Type	Description
sequence_id	string	Sequence prefix, e.g. 0826_0004
scene_type	string	materials or table_with_robot
tags	list	["test"] on video; ["test", "reconstruction"] on 3D slice

The import scripts also set source_tar and frame_count on the video slice when samples are built.

Frame-level fields (video slice only)

Each video sample has 121 frame documents (6,171 total across the dataset).

Field	FiftyOne type	Description
frame_id	int	Frame index from source metadata
sequence_id	string	Sequence id for this frame
max_depth	float	Depth scale factor from depth.json
depth_map	Heatmap	16-bit depth PNG on disk (range=[0, 65535])
normal_map	Heatmap	RGB-encoded normal PNG on disk
depth_json	dict	Parsed contents of depth.json
camera_extrinsics	list	4×4 extrinsics matrix
camera_intrinsics	list	3×3 normalized intrinsics matrix

There are no object detection or segmentation labels. Supervision is dense per-pixel depth, normals, and camera calibration.

---

3D Reconstruction

Each sequence has one merged point cloud for the whole scene (not one cloud per frame).

Process (reconstruct_scenes.py):

1. Decode 16-bit depth using paired depth.json.
2. Convert normalized intrinsics to pixel-space K.
3. Back-project depth pixels into world coordinates using per-frame extrinsics.
4. Color each 3D point from the matching RGB video frame.
5. Merge all 121 frames into one point cloud.
6. Downsample with Open3D voxel grid (voxel_size=0.01 by default).
7. Write scene.pcd and wrap it in scene.fo3d for the reconstruction group slice.

Default reconstruction settings:

Parameter	Default	Purpose
--stride	4	Subsample every 4th pixel during back-projection
--voxel-size	0.01	World-space voxel downsampling
--max-depth-ratio	0.999	Drop saturated far-plane depth values

The reconstruction is a point cloud, not a mesh. It is intended for interactive 3D viewing in FiftyOne, not watertight surface reconstruction.

---

How to Build / Reload the Dataset

Prerequisites

Download the selected source tars from Daniellesry/TransPhy3D

Run the two scripts in this order:

Step 1 — Extract assets and build video samples

python import_transphy3d.py --overwrite

This:

• Reads WebDataset tars from data/tars/test/
• Writes processed PNG/JSON assets and rgb.mp4 files
• Creates FiftyOne video samples with per-frame heatmaps and metadata

Step 2 — Reconstruct 3D scenes and build grouped dataset

python reconstruct_scenes.py --build-dataset --overwrite

This:

• Builds merged RGB point clouds (scene.pcd, scene.fo3d)
• Replaces/creates the grouped FiftyOne dataset TransPhy3D with video + reconstruction slices

---

Citation

If you use the source TransPhy3D dataset, cite:

@article{dkt2025,
  title   = {Diffusion Knows Transparency: Repurposing Video Diffusion for Transparent Object Depth and Normal Estimation},
  author  = {Shaocong Xu and Songlin Wei and Qizhe Wei and Zheng Geng and Hong Li and Licheng Shen and Qianpu Sun and Shu Han and Bin Ma and Bohan Li and Chongjie Ye and Yuhang Zheng and Nan Wang and Saining Zhang and Hao Zhao},
  journal = {https://arxiv.org/abs/2512.23705},
  year    = {2025}
}