--- license: cc-by-nc-nd-4.0 language: - en pretty_name: TEDWB1k size_categories: - 1K- TEDWB1k is derived from TED talks (https://www.ted.com), which are licensed under CC-BY-NC-ND 4.0 by TED Conferences, LLC. The dataset author claims no rights over the underlying TED content; this dataset is distributed for non-commercial academic research only and may be removed at any time at TED's request. By accessing TEDWB1k you agree to the terms below. extra_gated_fields: Name: text Email: text Affiliation: text Country: country Intended use: type: select options: - Academic research (university / lab) - Independent / personal research - Other (explain in 'Notes') Notes: text I will use TEDWB1k only for non-commercial academic research: checkbox I will credit TED in any publication, demo, or derivative work that uses TEDWB1k: checkbox I will not redistribute the data (in full or in part) outside my research group: checkbox I acknowledge that TED Conferences LLC owns the underlying video content and may request removal at any time: checkbox extra_gated_button_content: I agree β€” request access configs: - config_name: subjects data_files: - split: train path: metadata/subjects_train.parquet - split: train_subset_x1 path: metadata/subjects_train_subset_x1.parquet - split: train_subset_x12 path: metadata/subjects_train_subset_x12.parquet - split: train_val path: metadata/subjects_train_val.parquet - split: test path: metadata/subjects_test.parquet --- # TEDWB1k > πŸ‘€ **Want to browse before requesting access?** A public preview lives at > [`initialneil/TEDWB1k-preview`](https://huggingface.co/datasets/initialneil/TEDWB1k-preview). > It has the same 5 split tabs in the HF Dataset Viewer with thumbnails for all > 1,431 subjects, plus full downloadable data for 12 sample subjects β€” no > agreement required. Use that to explore. Use **this** (gated) repo when you > want the full training data. **1,431 TED-talk speaker videos with per-frame SMPL-X + FLAME tracking, ready for 3D human / avatar research.** TEDWB1k is a tracked subset of TED talks built with the [HolisticTracker](https://github.com/initialneil/HolisticTracker) (`ehm-tracker`) pipeline. Every subject is shot-segmented, background-matted, and fitted with whole-body SMPL-X (body + hands), FLAME (face + jaw + eyes), and per-shot whole-image keypoints. It is the dataset used to train the [HolisticAvatar](https://github.com/initialneil/HolisticAvatar) feed-forward Gaussian avatar model. > **License:** TED talks on ted.com are CC-BY-NC-ND 4.0. This dataset matches the upstream license: **CC-BY-NC-ND 4.0**. Non-commercial research only, attribution required, **no redistribution of modified or derivative versions**. ## At a glance | Split | Subjects | Approx download | Notes | |---|---:|---:|---| | `train_subset_x1` | 1 | ~80 MB | tiny single-subject overfit (βŠ‚ `train`) | | `train_subset_x12` | 12 | ~1 GB | 12-subject overfit (βŠ‚ `train`) | | `train_val` | 20 | ~2 GB | monitored during training (βŠ‚ `train`) | | `test` | 70 | ~10 GB | identity-disjoint evaluation set | | `train` | 1,361 | ~190 GB | full training pool | | **total** | **1,431** | ~200 GB | | `train` (1,361) and `test` (70) are identity-disjoint and together cover all 1,431 subjects. `train_subset_x1`, `train_subset_x12`, and `train_val` are all subsets of `train` β€” `train_val` are the 20 subjects whose frames the original training run reserved for validation monitoring (see `dataset_frames.json`); the small overfit subsets are intended for debugging. The HF Dataset Viewer above renders one row per subject with a thumbnail of the final tracked SMPL-X overlay (`track_smplx.jpg`) and the per-subject frame and shot counts. Switch between the 5 split tabs to browse each subset. ## Quick start ```bash pip install huggingface_hub # Smallest possible test (1 subject): python load_tedwb1k.py --split train_subset_x1 --out ./tedwb1k_x1 # 12-subject overfit set: python load_tedwb1k.py --split train_subset_x12 --out ./tedwb1k_x12 # 20-subject training-monitor set (subset of train): python load_tedwb1k.py --split train_val --out ./tedwb1k_train_val # 70-subject test set: python load_tedwb1k.py --split test --out ./tedwb1k_test # Full training pool (1361 subjects): python load_tedwb1k.py --split train --out ./tedwb1k_train ``` `load_tedwb1k.py` is included in this repo (or grab it from `ehm-tracker/release/load_tedwb1k.py`). It downloads only the matching subjects, merges per-subject tracking pickles into the format HolisticAvatar's `TrackedData` expects, extracts frames + mattes, and writes a fresh `extra_info.json` with absolute paths to the user's local data dir. After it finishes, point your training config at `--out`: ```yaml DATASET: data_path: ./tedwb1k_test ``` …and you can train / fine-tune HolisticAvatar with **zero code changes** to that codebase. ## Repository layout ``` TEDWB1k/ β”œβ”€β”€ README.md this file β”œβ”€β”€ train.txt 1,361 subject ids (full training pool) β”œβ”€β”€ train_subset_x1.txt 1 subject id (single-subject overfit, βŠ‚ train) β”œβ”€β”€ train_subset_x12.txt 12 subject ids (small overfit, βŠ‚ train) β”œβ”€β”€ train_val.txt 20 subject ids (training monitor, βŠ‚ train) β”œβ”€β”€ test.txt 70 subject ids (identity-disjoint evaluation) β”œβ”€β”€ dataset_frames.json frame-level train/valid/test split used by HolisticAvatar β”œβ”€β”€ metadata/ β”‚ β”œβ”€β”€ subjects_train.parquet per-split subject tables w/ embedded source-frame previews (HF Viewer) β”‚ β”œβ”€β”€ subjects_train_subset_x1.parquet β”‚ β”œβ”€β”€ subjects_train_subset_x12.parquet β”‚ β”œβ”€β”€ subjects_train_val.parquet β”‚ β”œβ”€β”€ subjects_test.parquet β”‚ β”œβ”€β”€ subjects.csv all rows in one CSV (programmatic use) β”‚ β”œβ”€β”€ skipped.txt (empty for the public release) β”‚ β”œβ”€β”€ previews/.jpg 1024Γ—1024 first source frame per subject (also embedded in parquets) β”‚ β”œβ”€β”€ ehm/.jpg full-res SMPL-X overlay grid (final tracking stage, ~13 MB) β”‚ β”œβ”€β”€ flame/.jpg full-res FLAME overlay grid (intermediate stage, ~6 MB) β”‚ └── base/.jpg full-res PIXIE+Sapiens overlay grid (stage 1, ~4 MB) └── subjects// β”œβ”€β”€ tracking/ β”‚ β”œβ”€β”€ optim_tracking_ehm.pkl per-frame SMPL-X + FLAME parameters β”‚ β”œβ”€β”€ id_share_params.pkl per-video shape / scale / joint offsets β”‚ └── videos_info.json frame-key listing for this video β”œβ”€β”€ frames.tar per-shot RGB JPGs (no audio, no video) └── mattes.tar per-shot RMBG-v2 alpha mattes ``` **Per-subject visualizations**: each subject has 4 standalone files under `metadata/`: - `metadata/previews/.jpg` β€” a clean 1024Γ—1024 source frame (the first frame of the first shot). These are what the HF Dataset Viewer renders in the `preview` column of the per-split parquets. - `metadata/ehm/.jpg` β€” the full-resolution SMPL-X overlay grid from the final tracking stage (large vertical contact sheet). - `metadata/flame/.jpg` β€” the FLAME overlay grid from the intermediate face-fitting stage. - `metadata/base/.jpg` β€” the stage-1 PIXIE+Sapiens overlay grid. You can fetch a single subject's QC visualizations without downloading the heavy `frames.tar`/`mattes.tar` by hitting any of those paths directly via `huggingface_hub.hf_hub_download`. Each `frames.tar` unpacks to: ``` /000000.jpg, 000001.jpg, ..., 0000NN.jpg ``` where `` is `NNNNNN_NNNNNN` encoding `start_frame_end_frame` β€” the inclusive keyframe indices of the shot inside the source TED talk, sampled at **0.5 fps** (one keyframe every 2 seconds). For example `000015_000019` is keyframes 15..19 (5 frames, covering seconds 30..38 in the source video). The JPGs inside the directory are indexed locally per shot starting at `000000.jpg`. We do **not** redistribute the original `.mp4` clips or the per-shot `audio.wav` extracts β€” those are direct excerpts of TED's source content. Only the per-frame JPGs (plus their alpha mattes) and the SMPL-X / FLAME tracking parameters are included. ## Tracking format Per-frame data inside `optim_tracking_ehm.pkl` (after the loader merge, keyed by `{video_id: {frame_key: ...}}`): ```python { 'smplx_coeffs': { 'global_pose': (3,), # axis-angle 'body_pose': (21, 3), # axis-angle per joint 'left_hand_pose': (15, 3), 'right_hand_pose': (15, 3), 'exp': (50,), # SMPL-X expression 'body_cam': (3,), 'camera_RT_params': (3, 4), }, 'flame_coeffs': { 'pose_params': (3,), 'jaw_params': (3,), 'neck_pose_params': (3,), # all zero (not optimized) 'eye_pose_params': (6,), # optimized 'eyelid_params': (2,), 'expression_params': (50,), 'cam': (3,), 'camera_RT_params': (3, 4), }, 'body_crop': {'M_o2c': (3,3), 'M_c2o': (3,3), ...}, 'head_crop': {'M_o2c': (3,3), 'M_c2o': (3,3)}, 'left_hand_crop': {...}, 'right_hand_crop': {...}, 'body_lmk_rlt': {'keypoints': (133,2), 'scores': (133,)}, 'dwpose_raw': {'keypoints': (133,2), 'scores': (133,), 'bbox': (4,)}, 'head_lmk_203': {...}, 'head_lmk_70': {...}, 'head_lmk_mp': {...}, 'left_mano_coeffs': {...}, 'right_mano_coeffs': {...}, } ``` Per-video identity data inside `id_share_params.pkl` (keyed by `{video_id: ...}`): ```python { 'smplx_shape': (1, 200), 'flame_shape': (1, 300), 'left_mano_shape': (1, 10), 'right_mano_shape': (1, 10), 'head_scale': (1, 3), 'hand_scale': (1, 3), 'joints_offset': (1, 55, 3), } ``` ## Pipeline Tracking was produced by `ehm-tracker` (a fork of LHM_Track) in three stages: 1. **`track_base`** β€” per-frame perception: - **PIXIE** for SMPL-X body initialization - **Sapiens 1B** for 133 whole-body keypoints - **HaMeR** for per-hand MANO regression on left/right hand crops - **MediaPipe** FaceMesh for 478-point face landmarks - additional 70- and 203-point face landmark models for face fitting - face / hand crop computation from the keypoints 2. **`flame`** β€” 2-stage FLAME optimization for face, jaw, expression, eyes, eyelids. 3. **`smplx`** β€” 2-stage whole-body SMPL-X optimization (body, hands, expression) consistent with the FLAME face fit. Each stage produces a sanity-check overlay grid (`track_base.jpg`, `track_flame.jpg`, `track_smplx.jpg`) that you can browse via the HF Dataset Viewer thumbnail column or in the per-subject directory. ## Known issues / caveats Please read these before training β€” they affect what is and isn't reliable in the data. - **`neck_pose_params` is all zero.** Not optimized by the pipeline; relying on neck rotation from FLAME will give you a static neck. - **Eyes only live in `flame_coeffs`.** `smplx_coeffs` has no `eye_pose` field β€” the `flame_coeffs.eye_pose_params` is the source of truth. They are non-zero (range roughly `[-0.54, 0.53]`). - **Per-subject pickles are flat.** If you skip the loader and read `subjects//tracking/optim_tracking_ehm.pkl` directly, the top-level keys are frame keys (e.g. `'000015_000019/000000'`), NOT video ids. The loader wraps them under `{video_id: ...}` so the merged file matches the format HolisticAvatar's `dataset/data_loader.py::TrackedData` expects. - **`dataset_frames.json` train/valid/test split is shot-limited.** During the original training run we limited val and test to the first 2 shots of each video to keep evaluation fast. The per-subject `videos_info.json` retains every shot, so the per-subject `optim_tracking_ehm.pkl` has all frames β€” only the merged `dataset_frames.json` is restricted. - **No videos, no audio.** We do not redistribute the original TED `.mp4` clips or the per-shot `audio.wav` extracts. Only per-frame JPGs (and their alpha mattes) plus the SMPL-X / FLAME tracking parameters are shipped. - **The original frames are stored as JPG** at the source resolution from `yt-dlp` of the TED talks. We did not re-encode. ## License [**CC-BY-NC-ND 4.0**](https://creativecommons.org/licenses/by-nc-nd/4.0/). Non-commercial research use only. Attribution required. **No derivatives** β€” you may not distribute modified or remixed versions of this dataset. The tracking parameters, JPG frames, and mattes are all derived works of TED talk videos that are themselves CC-BY-NC-ND on ted.com. This dataset matches the upstream license to remain compatible with TED's source restrictions. ## Links - Tracking pipeline: - HolisticAvatar (downstream model): - HF dataset: