Datasets:
|
Download README.md from KosmoResearch/GFM4DBenchMark: direct link, hf CLI and curl.
- Browser
- Download file 2.73 kB
-
https://huggingface.co/datasets/KosmoResearch/GFM4DBenchMark/resolve/main/README.md
- Command line
-
hf download hf://datasets/KosmoResearch/GFM4DBenchMark/README.md
-
curl -L -o README.md https://huggingface.co/datasets/KosmoResearch/GFM4DBenchMark/resolve/main/README.md
2.73 kB
| pretty_name: GFM4DBench | |
| tags: | |
| - computer-vision | |
| - 3d | |
| - video | |
| - point-tracking | |
| - 4d-reconstruction | |
| - benchmark | |
| # GFM4DBench data package | |
| [Paper](https://arxiv.org/abs/2611.16840) · [Project](https://kosmoresearch.github.io/Kosmo-GFM/) · [Code](https://github.com/KosmoResearch/Kosmo-GFM) · [Model](https://huggingface.co/KosmoResearch/Kosmo-GFM) | |
| TapVid3D / WorldTrack mini and SynthVerse benchmark data used by GFM4DBench. | |
| ## Layout | |
| ```text | |
| 4dbenchmark_data/ | |
| ├── Tapvid3d_mini/ | |
| │ └── Tapvid3d_mini_extracted/ | |
| │ ├── adt_mini.json | |
| │ ├── pstudio_mini.json | |
| │ ├── po_mini.json | |
| │ ├── ds_mini.json | |
| │ ├── adt_mini/ | |
| │ ├── pstudio_mini/ | |
| │ ├── po_mini/ | |
| │ └── ds_mini/ | |
| └── SynthVerse/ | |
| ├── synthverse_mini.json | |
| └── converted/ | |
| ├── converted_subset50_64f_dynamic/ # trajectory arrays | |
| │ └── <sequence>/{trajs_3d,visibs} | |
| └── <sequence>/rgbs/ | |
| ``` | |
| Each TapVid3D subset has 50 sequences. Every sequence contains the RGB frames | |
| and the per-frame 3D trajectories and visibility flags. | |
| SynthVerse has 49 sequences of 64 frames and uses the PointOdyssey coordinate | |
| convention. It is reported separately from the TapVid3D average. | |
| ## Usage | |
| Point `GFM4D_DATA_ROOT` at this directory (the one containing | |
| `Tapvid3d_mini/` and `SynthVerse/`) and run the GFM4DBench evaluator: | |
| ```bash | |
| export GFM4D_DATA_ROOT=/path/to/4dbenchmark_data | |
| bash bench/GFM4DBench/script/eval.sh dynamic-points --subset pstudio --limit-seqs 1 | |
| bash bench/GFM4DBench/script/eval.sh dynamic-points --subset synthverse | |
| ``` | |
| The registry shipped with GFM4DBench (`bench/GFM4DBench/configs/subsets.yaml`) | |
| already uses the nested paths above. | |
| ## Expected query counts | |
| With the default clip length of 64 frames: | |
| | Subset | Sequences | Frames per sequence | Scored queries | | |
| |---|---|---|---| | |
| | `adt` | 50 | 300 | 22187 | | |
| | `pstudio` | 50 | 150 | 8720 | | |
| | `po` | 50 | 128 | 53465 | | |
| | `ds` | 50 | 128 | 45149 | | |
| | `synthverse` | 49 | 64 | 15762 | | |
| `total_queries` in `summary.json` must match these values. | |
| ## Contents and exclusions | |
| Included: `rgb`, `trajs_3d` and `visibs`; intrinsics and extrinsics are stored | |
| directly in the JSON. | |
| Excluded: `depth/`, `trajs_2d/`, `valids/`, `query_track_indices/`, `*.bak` | |
| duplicates and the `adt_mini_trajs2d_fixed/` copy. The evaluator never reads | |
| them. The JSON files have been cleaned so that no path points to a missing | |
| file, and the SynthVerse RGB files are stored as regular files rather than | |
| symlinks. | |
| ## Integrity | |
| ```bash | |
| sha256sum -c SHA256SUMS | |
| ``` | |
| `MANIFEST.json` records the subset statistics and the layout version. | |