Analysis of Features-Brightness Bias in Computer Vision

This paper explores the impact of mask brightness on various computer vision techniques. We conducted experiments on multiple applications, including reconstruction, generation, object recognition, and vectorization. Darker features relative to the background resulted in improved performance for autoencoders, Potrace vectorization, and GAN training stability. However, this relationship did not hold for Mask R-CNN and YOLO. Nonetheless, the brightness difference between masks and backgrounds influenced object recognition confidence. Our findings highlight the importance of considering brightness variations to avoid biases and provide insights for further research