- Rating
- 3
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes a novel CCSG-WB framework to address the problem of sRGB WB correction: how to disentangle intrinsic scene colors from color casts. Owing to the strong non-linear operations of the camera ISP pipeline, previous works fail to handle cases when the dominant colors of the scene are similar to the color cast. To resolve this ambiguity, this paper introduce an image-level representation characterizing cast tendency called Color Cast State (). The framework operates in two stage. In Stage I, CCSRL uses paired images and a soft similarity comparison learning setup. It trains a CCS encoder to learn a well-structured CCS space. In Stage II, the setup uses the frozen CCS encoder to extract . It then adjusts a U-shaped Transformer backbone through CCSG blocks. These blocks use FiLM parameters and . The Rendered WB dataset and the Rendered Cube+ dataset show that the method gives the best known accuracy. They also show that the method works well across cameras. They further show that the method has very low delay.
Critical strengths
- Well-Motivated Problem Formulation: The paper pinpoints the core ambiguity of sRGB-WB correction: the dominant intrinsic color of the scene, after the non-linear ISP pipeline, looks like a color cast. Modeling the color cast explicitly as a conditional guidance vector is an insightful approach.
- Sound Methodological Design: The method combines a tendency loss , a strength loss , and a soft comparison loss . This combination effectively avoids hard label edges. This combination builds a continuous and smooth color tint space.
- Strong Empirical Results: The proposed method consistently achieves the best performance compared to state-of-the-art baselines (e.g., ABC-Former, DCLAN) for the MSE, MAE, and metrics on the Rendered WB and Rendered Cube+ datasets.
- Efficiency and Robustness: The method takes about per image on the Set 1-Test dataset. It gets the lowest across all color temperatures between 2850K and 7500K. It demonstrates robust adaptation of casting.
Critical weaknesses
-
Stage Ⅰ of the model strongly relies on pairs of images with incorrect white balance and their corresponding ground truth, which is impractical for most unpaired and in-the-wild images, where rendering paired images is impossible.
-
The experiments are restricted to global and uniform color casts (2850K-7500K). The method is not evaluated or discussed on scenes with mixed illumination and spatially varying white balance.
-
The running speed remains competitive. However, the CCS encoder is built on a pre-trained ViT-B model. The method resizes input images to during feature extraction. However, the method still keeps a large ViT next to a Transformer repair network. This design may cause problems when people use it on devices that have limited resources.
Questions for authors
- How does the global representation perform on images with spatially non-uniform illumination or mixed light sources (e.g., warm light on the left and cool light on the right of the scene)?
- The CCS encoder relies on a pre-trained ViT-B/32 fine-tuned with LoRA. Could a lighter CNN or compact Transformer backbone be substituted to reduce parameter overhead? And how much performance degradation in would result?
- In severely under-exposed or over-exposed regions where nonlinear ISP processing leads to hue clipping and detail degradation, how does the global FiLM modulation (, ) avoid introducing color distortion and visual artifacts?
Response to Reviewer vfug
Thank you for the thorough review of our paper and the insightful comments.
W1. Paired Supervision and Real-World Applicability
The paired incorrectly white-balanced and ground-truth images serve only as supervision signals for training stage and are not required during inference. After CCS representation learning is completed, the whole framework takes only a single sRGB image as input at inference time, and the trained CCS encoder directly captures the color cast state from the input.
Furthermore, Rendered WB Set 2 and Rendered Cube+ are exclusively used for testing and excluded from training. Their image formation pipeline closely resembles the process of erroneous on-camera white balance. This benchmark provides a controlled and practically relevant setting for evaluating the WB correction. Our method achieves a mean ΔE2000 of 4.02 on Rendered WB Set 2 and 3.68 on Rendered Cube+. These results demonstrate that our CCSG-WB generalizes to real-world scenes. When in-the-wild retraining is desired, paired training data can be constructed following KNNWB [1], as long as RAW images are accessible.
W2/Q1. Global CCS guidance and spatially varying illumination
Our CCS is designed to characterize chromatic deviations arising from global white balance errors, help distinguish color casts from intrinsic scene colors, and deliver correction guidance for the shared correction backbone. The current formulation mainly targets global color casts, consistent with the evaluation settings of Rendered WB and Rendered Cube+. WB correction under mixed illumination and spatially varying lighting constitutes a related yet distinct task. We will plan to explore spatially varying or region-wise CCS conditioning as a future extension.
W3/Q2. Lightweight CCS Encoder and Accuracy–Efficiency Trade-off
Reducing the computational overhead of the CCS encoder is critical for practical deployment. To this end, we only replace the CCS encoder with a lightweight ResNet-18 encoder while keeping all other settings unchanged.
The experimental results show that after replacing the CCS encoder with this lightweight CNN encoder, the mean ΔE2000 is 4.01, an absolute gap of 0.33 compared with the ViT-B/32 variant (3.68).
Q3. Color Fidelity under Challenging Exposure Conditions
In our method, the global CCS representation modulates intermediate features via FiLM to provide guidance. The final correction result is jointly determined by the global color cast state and local image features. Furthermore, the correction backbone integrates multi-scale feature extraction, attention mechanisms, and skip connections to preserve local image information during reconstruction. Combined with the reconstruction loss, these components enable the network to correct color casts while reducing undesired color distortions and visual artifacts. We will add visual results under challenging exposure conditions to the supplementary material in the revised manuscript.
[1] When color constancy goes wrong: Correcting improperly white-balanced images. CVPR 2019.