Cui Z, Harada T. RAW-adapter: Adapting pre-trained visual model to camera RAW images[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 37-56.
Background
- RAW images:
- directly acquired by the camera sensor--encompasses abundant information.
- sRGB images:
- derived from RAW data through image signal processor (ISP)--ease of acquisition and efficient storage.
Existing methods
- Fig. 1(a) ISP process
- Fig. 1(b) manually designed ISP--based on visual experience
- Fig. 1(c) optimize ISP jointly with downstream networks
- Fig. 1(d) RAW-Adapter
Introduction
Method
- RAW-Adapter--adapting sRGB pre-trained models to camera RAW images
- input-level adapters: query adaptive learning (QAL) and implicit neural representations (INR) optimize ISP key parameters
- model-level adapters: prior input stage information enriches model understanding
Method--input-level adapters
- Gain & Denoise
- White Balance & CCM Matrix
- Color Manipulation Process
Method--model-level adapters
- employ the prior information from ISP process as model-level adapters to guide subsequent models' perception
Evaluation
- datasets: PASCAL RAW [1] and LOD [2]
1\] Omid-Zohoor A, Ta D, Murmann B. PASCALRAW: raw image database for object detection\[J\]. Stanford Digital Repository, 2014, 2(3): 4. \[2\] Hong Y, Wei K, Chen L, et al. Crafting Object Detection in Very Low Light\[C\]//BMVC. 2021, 1(2): 3.      *** ** * ** ***