NeurIPS 2020 Highlights Novel Approaches in Computer Vision Research
Leading tech labs present innovative computer vision papers at NeurIPS 2020, focusing on segmentation, object detection, and texture bias. One standout project uses causal inference to significantly improve weakly-supervised semantic segmentation.
NeurIPS 2020 showcases a wide array of innovative computer vision research papers from leading labs like Google Brain, MIT, and Adobe Research. These studies introduce novel approaches to semantic and instance segmentation, object detection, and image manipulation using autoencoders. Additionally, researchers actively investigate texture bias in convolutional neural networks and the underlying qualities of adversarially robust models.
One standout paper from Nanjing University of Science and Technology and Alibaba Group tackles weakly-supervised semantic segmentation using a causal inference framework. The researchers address the challenge of generating accurate pixel-level pseudo-masks using only image-level labels. They identify that ambiguous boundaries in these masks stem from confounding contexts, such as the co-occurrence of objects like a horse and a person, which mislead standard model inspections.
To resolve this issue, the team develops a method called Context Adjustment (CONTA) based on a structural causal model. This technique effectively removes confounding bias from image-level classification to produce clearer pseudo-masks. These refined masks then serve as highly accurate ground truth for the subsequent segmentation model, ultimately improving overall performance on benchmark datasets like PASCAL VOC 2012.