Exploring Color Invariance through Image-Level Ensemble Learning

TL;DR

This work tackles color bias caused by lighting and camera variation in wide-area surveillance tasks such as person re-identification and industrial dust segmentation. It introduces Random Color Erasing (RCE), an image-level ensemble learning strategy that combines global grayscale transformations with local grayscale patches to rebalance color features and non-color discriminative cues; the method uses probabilities $p_g$ and $p_r$ to control global and local erasing and can be viewed as an efficient alternative to GAN-based style transfer. The authors provide theoretical insights suggesting that ensembling grayscale- and color-trained components can reduce generalization error, and they validate the approach through extensive experiments across two tasks, multiple datasets, and several strong baselines, with notable improvements in cross-domain settings. Visual analyses with Grad-CAM support the claim that RCE yields more robust attention to truly discriminative regions under color variations, underscoring practical impact for robust color-invariant vision systems.

Abstract

In the field of computer vision, the persistent presence of color bias, resulting from fluctuations in real-world lighting and camera conditions, presents a substantial challenge to the robustness of models. This issue is particularly pronounced in complex wide-area surveillance scenarios, such as person re-identification and industrial dust segmentation, where models often experience a decline in performance due to overfitting on color information during training, given the presence of environmental variations. Consequently, there is a need to effectively adapt models to cope with the complexities of camera conditions. To address this challenge, this study introduces a learning strategy named Random Color Erasing, which draws inspiration from ensemble learning. This strategy selectively erases partial or complete color information in the training data without disrupting the original image structure, thereby achieving a balanced weighting of color features and other features within the neural network. This approach mitigates the risk of overfitting and enhances the model's ability to handle color variation, thereby improving its overall robustness. The approach we propose serves as an ensemble learning strategy, characterized by robust interpretability. A comprehensive analysis of this methodology is presented in this paper. Across various tasks such as person re-identification and semantic segmentation, our approach consistently improves strong baseline methods. Notably, in comparison to existing methods that prioritize color robustness, our strategy significantly enhances performance in cross-domain scenarios. The code available at \url{https://github.com/layumi/Person\_reID\_baseline\_pytorch/blob/master/random\_erasing.py} or \url{https://github.com/finger-monkey/Data-Augmentation}.

Figures (5)

• Figure 1: The retrieval results of the model trained with visible (RGB) image and the model trained with grayscale image on the Market1501 dataset. The numbers on the images indicate the rank of similarity in the retrieval results, the red and green numbers denote the wrong and correct results, respectively.
• Figure 2: The figure illustrates two approaches for learning color invariance. The left image demonstrates the integrated model trained with samples containing and lacking color information separately. On the right, our proposed image-level ensemble learning strategy, Random Color Erasing, achieves color invariance by applying random local homogeneous grayscale transformations to integrate samples with and without color information simultaneously, thereby achieving equivalent effects to ensemble models
• Figure 3: Analysis of Random Color Erasing Strategy. Assuming that the prediction of component networks are combined by majority voting, each component network is regard as a classifier. The component predictions are combined via weighted voting for classification, where the weights are determined by the algorithms themselves. Some of classifiers are trained using visible images, while others are trained using grayscale images.
• Figure 4: The performance evaluation of our Global Color Erasing (GCE) and our Random Color Erasing (RCE) involves varying hyperparameters on the Market1501 dataset of ReID . Our evaluation results are presented with the default usage of re-ranking.
• Figure 5: Comparison of Grad-CAM activation maps between a normally trained model and a model trained using our Random Color Erasing.