The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
EgoTouch clean atomic visual feature cache v1
Frozen RGB feature cache for the P4 VT2P, VP2T and TP2V experiments in
kaichen-z/VLA-HAND@8acafb5.
It must be paired with
egotouch-clean-atomic-cache-v3.
Features were extracted from the published chest.mp4 streams by each atomic
frame's video_decode_frame. The source RGB is available from
zhouzhoujy/EgoTouch
and is not duplicated here.
Feature contract
- Backbone: OpenTouch ViT-B/16 visual encoder.
- Source checkpoint:
opentouch-vp2t-encoder-best/epoch_280.pt. - Checkpoint SHA-256:
5b8d3ce311bad8f50515e2c4b72a054283dc1cce67ef82696bb74c26f9b062d8. - Image size:
224 x 224; ImageNet mean/std normalization. - Stored feature: 768D CLS token before the learned 64D projection.
- Temporal aggregation: mean over the eight history frames during training.
- Coverage: 1,534 recordings, 1,899,635 unique decoded RGB frames and 3,492,441 atomic-frame placements.
Each split contains:
visual_cls.npy float16 [atomic_frames, 768]
visual_valid.npy bool [atomic_frames]
video_frame.npy source decoded frame index
cache_manifest.json pins the paired cache-v3 episode/frame ordering, while
visual_manifest.json records extraction provenance. Copy or link these three
arrays into the corresponding split of the tactile cache before using the
current egotouch_vt2p dataloader.
- Downloads last month
- 35