nguyenminh04/coin-feats

COIN: videos and InternVideo2 + BEATs features (private, work in progress)

COIN (Tang et al., CVPR 2019) contains 11,827 YouTube instructional videos covering 180 tasks in 12 domains, for 468 hours in total. It has 46,354 step segments with timestamps and a step label, for example remove the old memory chip. This repo holds the videos we could still download in 2026 and one feature vector per second. The features are extracted exactly like our YouCook2 features (nguyenminh04/uniav-youcook2-data, iv2_feats/), so the two datasets can be trained together.

Code: tools/coin_hub.py and tools/extract_internvideo2.py on branch coin-extract of im-xiaoming/UniAV-fixed. Status (2026-10-06): downloading is stopped and 5,180 videos have features (63 shards: 0000-0015 from the PC's first downloads, 1000-1099 from the distributed download, 2000-2001 from a retry with login cookies; 3,947 from the COIN training split, 1,233 from the testing split; about 20,000 step segments, 743 distinct labels). Not downloaded: about 4,400 videos of chunks 1031-1085 and 38 videos that need an age-verified or member account. See videos/manifest.json, dl/done/ and the feats/ folder.

Layout

Path	Content
feats/coin_feats_NNNN.tar	features: one <youtube_id>.npz per video of shard NNNN, plus failed.txt (videos the extractor could not read; usually empty)
videos/coin_videos_NNNN.tar	100 source videos per shard, <youtube_id>.mp4, at most 480p, video + audio (yt-dlp)
videos/manifest.json	{"NNNN": [youtube_id, ...]}: which videos are in which shard
videos/DONE	written when downloading stopped (2026-10-06)
exclude.txt	COIN video ids not to train on: 3 of them are YouCook2 validation videos too (jT75QMjRkD0, e1gtgMczUwE, LWuuCndtJr0); training on them would inflate YouCook2 val scores
dl/	bookkeeping of the distributed download: plan.json, plan_retry.json, done/CCCC.json (ids downloaded, failed, to retry), claims/
claims/NNNN__<worker>	bookkeeping: which GPU worker took which shard
text_feats/coin_label_emb_iv2.npz	InternVideo2 text vectors (512-d) of the 749 step labels: labels, emb (centred on the COIN label mean, pairwise cosine 0.01), raw (normalised, uncentred), mean, mean_youcook2. Same text tower as YouCook2's caption_emb_iv2.npz; made by tools/embed_captions.py --coin
parity/6uHoTJSLoL8.mp4	a YouCook2 video used to check that every worker reproduces the YouCook2 features (cosine 1.00000)

About 18% of COIN videos are no longer available on YouTube (deleted or private), so roughly 9,700 videos can be downloaded at all. Shard numbers are not contiguous (0000-0015, 1000-1099 where chunks 1031-1085 were never downloaded, 2000-2001). The annotations are not copied here. They come from coin-dataset/annotations (COIN.json), keyed by the same YouTube id.

Feature format

Each .npz holds three float16 arrays with one row per second of video, with row i centred on i + 0.5 s:

Key	Shape	What
v768	(T, 768)	InternVideo2-Stage2-1B attention-pooled vision embedding of 4 frames (2 fps, 224×224) around second i; raw norm ~50
v512	(T, 512)	vision_proj(v768), L2-normalised, in InternVideo2's video-text space (its BERT-large text tower embeds sentences in the same space)
a768	(T, 768)	BEATs (from InternVideo2-Stage2-6B-Audio, audio_6b.pth) on a 3 s window of 16 kHz mono audio centred on second i, mean-pooled; zeros if the video has no audio

Checkpoints: OpenGVLab/InternVideo2-Stage2_1B-224p-f4 (gated) and OpenGVLab/InternVideo2-Stage2-6B-Audio, fp16. Our models L2-normalise every row before use (iv2_l2norm: true).

Samples

The first videos of feats/coin_feats_0000.tar, with their COIN annotation:

youtube id	task (COIN class)	duration	rows	steps (seconds: label)
0R9pdc9dO3Q	ReplaceBatteryOnTVControl	122.4 s	123	14–16 open cover · 16.5–19 remove battery · 108–113.5 put battery in · 114–121.5 close cover
0xybXrdeSTk	CutMango	92.7 s	93	5–12 cut the wide side · 21–34 remove the core · 54–58 cut the wide side · 60–72 remove the core
1nQp_Yb8YW0	ReplaceMemoryChip	165.6 s	166	64–73 take out the shell · 74–90 remove the old memory chip · 95–125 install the new memory chip · 126–135 fit on the shell
2zqpJb-3SyI	CutGrapeFruit	174.9 s	175	126–128 cut in half · 131–133 slice the pulp

import json, tarfile, numpy as np
from huggingface_hub import hf_hub_download

tar = hf_hub_download('nguyenminh04/coin-data', 'feats/coin_feats_0000.tar', repo_type='dataset')
tarfile.open(tar).extractall('coin_feats')
z = np.load('coin_feats/0R9pdc9dO3Q.npz')
{k: z[k].shape for k in z.files}   # {'v768': (123, 768), 'v512': (123, 512), 'a768': (123, 768)}

coin = json.load(open('COIN.json'))['database']                    # from coin-dataset/annotations
for a in coin['0R9pdc9dO3Q']['annotation']:
    start, end = a['segment']                                      # seconds
    rows = z['v768'][int(start):int(end) + 1]                      # features of that step
    print(a['label'], rows.shape)

Differences from the YouCook2 features

• Same script, same settings, same checkpoints: every worker reproduces the stored YouCook2 features of 6uHoTJSLoL8 with cosine 1.00000 for all 182 rows (on Colab torch 2.11 and Kaggle torch 2.0).
• The source encode differs: YouTube serves a different encode than the one YouCook2 was downloaded from. Re-downloading a YouCook2 video with yt-dlp's default format gives video cosine ~0.997 (min 0.989) to the stored features, and audio is identical.
• COIN labels are short step names, not sentences: 749 distinct labels of 4.9 words on average, each reused across videos of the same task. YouCook2 captions average 8.8 words and are nearly all unique.

License

COIN annotations and videos belong to their authors and to YouTube uploaders: research use only, under COIN's terms. This repo is private and is not for redistribution.