Brief Review — Logarithmic Lenses: Exploring Log RGB Data for Image Classification
Psuedo Log RGB
Logarithmic Lenses: Exploring Log RGB Data for Image Classification
Logarithmic Lenses, by Northeastern University
2024 CVPR (Sik-Ho Tsang @ Medium)Image Classification
1989 … 2024 [FasterViT] [CAS-ViT] [TinySaver] [Fast Vision Transformer (FViT)] [MogaNet] [RDNet]
==== My Other Paper Readings Are Also Over Here ====
- Authors enhance the Log RGB approach in 2023 BMVC.
- It is found that networks trained on Log RGB data exhibit improved performance on an unmodified test set and invariance to intensity and color balance modifications without additional training or data augmentation.
- It is also found that the gains from using high quality Log RGB could also be partially or fully realized from data in 8-bit sRGB-JPG format by inverting the sRGB transform and taking the log.
Outline
- Log RGB
- Pseudo Log RGB
1. Log RGB
In the Log RGB approach in 2023 BMVC, with the new RAW10 dataset using RAW format, it is found that using Log RGB helps to boost the classification performance. (Please read this paper first before reading this for the motivation.)
A custom 11-layer CNN is designed. And it is also found that using Log RGB helps to boost the classification performance.
2. Pseudo Log RGB
Pseudo-linear and pseudo-log are introduced to designate data that was originally 8-bit sRGB and has had an inverse sRGB transformation applied (pseudo-linear) followed by a log transform (pseudo-log).
Using the pseudo-log data provided a benefit compared to a network trained on the original sRGB data, particularly for the modified test sets.
- Conversely, the differences between training on pseudo-log data or true-log data were small.
This finding is significant because it means that the millions of images in existing annotated data sets may be useful for training log of RGB networks, even if those networks are then used with true log data.
