Syn4D's picture
Document benchmark-contained evaluation asset layout
00b01d5 verified
|
Raw History Blame Contribute Delete
3.94 kB
---
pretty_name: Syn4D multi-task benchmark files
---
# Syn4D multi-task benchmark files
This directory contains the portable ground-truth release and protocol metadata
for the 512-sequence Syn4D benchmark. The evaluator and model adapters are
maintained in [Syn4D_benchmark](https://github.com/jzr99/Syn4D_benchmark).
## Contents
- `data/tracking_gt/<variant>/<scene>/<sequence>.npy`: one structured tracking
ground-truth file per sequence;
- `sequences.jsonl`: portable sequence index with paths relative to a local
`kaggle_eval` root;
- `protocol.json`: task definitions, alignment rules, metrics, and aggregation;
- `baselines.json`: official baseline repositories, revisions, and task support;
- `SHA256SUMS`: integrity checksums for all tracking ground-truth files.
There are 512 sequences: four variants (`mixed`, `mixed_no_bedlam`, `og`, and
`sim`) times eight scenes, eight base sequences, and two cameras. Every task
uses frames `0, 6, ..., 186` (32 frames).
## Tracking NPY schema
Each file is a non-pickled structured NumPy scalar using schema
`syn4d-tracking-gt-v1` and can be loaded with `np.load(path,
allow_pickle=False)`.
| Field | Shape | Description |
| --- | --- | --- |
| `schema` | scalar | Schema identifier |
| `sequence_id` | scalar | `<variant>/<scene>/<sequence>` |
| `tracks_ref0` | `[32,512,3]` | 3D trajectories in OpenCV frame-0 camera coordinates |
| `visibility` | `[32,512]` | Valid/visible correspondence mask |
| `dynamic` | `[512]` | Dynamic-query labels |
| `query_uv` | `[512,2]` | Frame-0 source-image query pixels |
| `frame_indices` | `[32]` | Evaluated source-frame indices |
| `intrinsics` | `[4]` | `[fx, fy, cx, cy]` |
| `image_size` | `[2]` | `[height, width]` |
The tracking benchmark evaluates these fixed 512 sparse queries. The
intermediate 4096-query WorldTrack conversion files are not required for
evaluation and are intentionally omitted.
Model inputs and the remaining task ground truth are included under this
directory's `challenge_eval` subtree:
- `<variant>/<scene>/mp4/<sequence>.mp4`: 512 model-input videos;
- `<variant>/<scene>/exr_layers/depth/<sequence>/`: the 32 evaluated depth EXRs;
- `<variant>/<scene>/ground_truth/meta_exr_csv/`: camera-pose metadata.
## Download and use
```bash
git clone https://github.com/jzr99/Syn4D_benchmark.git
cd Syn4D_benchmark
python -m pip install -r requirements.txt
python download_benchmark.py --output data/release
export SYN4D_DATA_ROOT="$PWD/data/release/benchmark/challenge_eval"
export SYN4D_TRACKING_GT="$PWD/data/release/benchmark/data/tracking_gt"
```
The downloader fetches only the release files needed for evaluation, resumes
partial downloads, and verifies the expected file counts. It does not download
the much larger source-PNG mirror. See the GitHub README for task-specific
downloads and the prediction contract.
Verify the release with:
```bash
cd Syn4D_Benchmark/benchmark
sha256sum --check SHA256SUMS
```
## Baseline results
All values are macro-averages over 512 sequences. A dash means the official
method does not expose that task through a compatible inference interface.
| Method | Track score | APD | Dynamic APD | EPE (m) | Depth AbsRel | Depth delta1 | Pose ATE (m) | RPE trans. (m) | RPE rot. (deg) |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| Open-D4RT | 0.3043 | 0.3631 | 0.2317 | 2.3298 | 0.1290 | 0.8408 | 0.0676 | 0.0355 | 0.3872 |
| 4RC | 0.4763 | 0.5848 | 0.3417 | 1.5925 | 0.0775 | 0.9195 | 0.0370 | 0.0218 | 0.1593 |
| V-DPM | 0.4880 | 0.5774 | 0.3757 | 1.6371 | 0.0891 | 0.9022 | 0.0375 | 0.0265 | 0.1693 |
| Any4D | 0.2776 | 0.3676 | 0.1701 | 1.8578 | 0.1040 | 0.8874 | 0.1611 | 0.1382 | 0.7645 |
| TraceAnything | 0.2206 | 0.3240 | 0.1003 | 2.3577 | - | - | - | - | - |
| St4RTrack | 0.2575 | 0.3302 | 0.1667 | 2.2390 | - | - | - | - | - |
| SpaTrackerV2 | 0.4334 | 0.5452 | 0.2960 | 1.6163 | 0.0727 | 0.9210 | 0.0444 | 0.0332 | 0.1569 |