docs: Compare augmentation pipelines between DALI dataloaders

[Copilot]

Comparison of data augmentation between ap_dataloader_dali.py and ap_dataloader_dali_codec.py.

Summary

Color/brightness augmentations are identical between both files:

• Brightness/Contrast: 80% prob, contrast (0.6, 1.4), brightness (-0.125, 0.125)
• Saturation: 80% prob, range [0.6, 1.4]
• Hue: 80% prob, range [-0.2, 0.2]
• Color Space (RGB→BGR): 10% prob

Key Differences

The distinction is in data source and processing, not augmentation:

Feature	dali.py	dali_codec.py
Frame sampling	Segment-based	Linear uniform (np.linspace)
Motion vectors	❌	✅ HEVC codec extraction
Residual features	❌	✅ Extracted and processed
visible_indices	❌	✅ Top-K patch selection
Caching	❌	✅ Caches visible_indices
MV compensation	N/A	none/median/mean/similarity/affine
Static video fallback	N/A	✅ Hybrid uniform + Top-K
Center prior	N/A	✅ Gaussian weighting

The codec version extracts motion vectors and residuals to compute importance-weighted visible_indices for attention-based models.

Original prompt

https://github.com/EvolvingLMMs-Lab/OneVision-Encoder/blob/main/dataloader/ap_dataloader_dali_codec.py

https://github.com/EvolvingLMMs-Lab/OneVision-Encoder/blob/main/dataloader/ap_dataloader_dali.py

帮我看看这两个aug的区别

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.