dmayboroda/tvals

Shot Embeddings

What was done

1. Identified videos with shot annotations

video_id	Shots
6436496163	3
7016927533	3
8070110346	1
911002795	1

2. Retrieved shot boundaries

shot_id	start (s)	end (s)	duration (s)
6436496163_0	0.00	67.60	67.60
6436496163_1	67.64	139.20	71.56
6436496163_2	139.24	189.32	50.08
7016927533_0	0.00	15.12	15.12
7016927533_1	15.16	49.36	34.20
7016927533_2	49.40	82.52	33.12
8070110346_0	0.00	47.92	47.92
911002795_0	0.00	118.52	118.52

3. Frame embeddings

Frame-level embeddings were queried from the S3 Tables Iceberg table via Athena:

• Model: ViT-B/32 (CLIP)
• Embedding dimensionality: 512
• Sample FPS: 2.0 (one frame every 0.5s)
• Total frames fetched for 4 videos: 884
• Frames that fell in gaps between shots: 5 (on boundary edges)

4. Grouped into per-shot .npy files

Each frame was assigned to a shot based on shot_start_timestamp <= frame_timestamp_seconds <= shot_end_timestamp. Embeddings within each shot are sorted by timestamp and stacked into a single numpy array.

Output files

File	Shape	Frames	Size
6436496163_0.npy	(136, 512)	136	272 KB
6436496163_1.npy	(143, 512)	143	286 KB
6436496163_2.npy	(100, 512)	100	200 KB
7016927533_0.npy	(31, 512)	31	62 KB
7016927533_1.npy	(68, 512)	68	136 KB
7016927533_2.npy	(67, 512)	67	134 KB
8070110346_0.npy	(96, 512)	96	192 KB
911002795_0.npy	(238, 512)	238	476 KB

File format

• Standard NumPy .npy format
• dtype: float32
• Shape: (num_frames, 512) — rows are frames sorted by timestamp, columns are ViT-B/32 embedding dimensions
• Naming: {video_id}/{video_id}_{shot_index}.npy

Loading

import numpy as np

embeddings = np.load("shot_embeddings/6436496163/6436496163_0.npy")
print(embeddings.shape)  # (136, 512)

# Mean embedding for the shot
shot_embedding = embeddings.mean(axis=0)  # (512,)