tools/taco_analysis/profile_comparison.md

TACO Dataset Profile Comparison: Original vs Resized

Generated: 2026-02-23 | Config: 4 ctx + 3 tgt views, 4 past + 8 future frames (52 decodes/sample) | 30 samples, no caching

Dataset Info

	Original (4096x3000)	Resized (512x376)	Speedup
Native resolution	4096x3000 (11 cams), 2048x1500 (1 cam)	512x376 (12 cams)	—
Total disk size	2.2 TB	495 GB	4.4x smaller
Allocentric videos	809 GB	4.3 GB	188x smaller
2D Segmentation	632 GB	145 GB	4.4x smaller
Sequences	2120	2120	same
Total samples	307,117	307,117	same

Single-Sample Loading (cold, no cache)

	Original	Resized	Speedup
Without segmentation
Mean	3.91 s	157.5 ms	24.8x
Median	3.85 s	159.0 ms	24.2x
P10	3.28 s	131.3 ms	25.0x
P90	4.74 s	188.7 ms	25.1x
With segmentation
Mean	4.23 s	169.3 ms	25.0x
Median	4.12 s	168.0 ms	24.5x
P10	3.41 s	125.8 ms	27.1x
P90	5.17 s	220.2 ms	23.5x

Batch Throughput (samples/sec)

Workers	Original	Resized	Speedup
0	0.27	5.70	21.1x
2	0.33	6.98	21.2x
4	0.52	12.22	23.5x

Per-Component Breakdown (mean, cold)

Component	Original	Resized	Speedup
Video decode (decord)	4.05 s	125.8 ms	32.2x
Frame resize (bilinear)	1.40 s	41.0 ms	34.1x
Camera params (JSON)	2.8 ms	1.5 ms	1.9x
Segmentation masks (npy)	147.8 ms	6.2 ms	23.8x
Full getitem	3.37 s	154.0 ms	21.9x

Key Takeaways

1. ~25x faster loading across the board — resized dataset loads a full sample in ~160ms vs ~4s
2. Video decode dominates in both cases (~80-120% of getitem time), but drops from 4s to 126ms
3. Resize cost nearly eliminated: 1.4s → 41ms (frames are already at target resolution, only minor interpolation needed)
4. Segmentation masks 24x faster: 148ms → 6ms (375×512 vs 750×1024, plus smaller file I/O)
5. 4.4x less disk space overall (2.2TB → 495GB), with allocentric videos 188x smaller (809GB → 4.3GB)
6. Throughput with 4 workers: 12.2 samples/sec (resized) vs 0.52 samples/sec (original)