Domain Decorrelation with Potential Energy Ranking

Machine learning systems, especially the methods based on deep learning, enjoy great success in modern computer vision tasks under ideal experimental settings. Generally, these classic are built i.i.d. assumption, supposing training and test data drawn from same distribution independently identically. However, aforementioned assumption is, general, unavailable real-world scenarios, as a result, leads to sharp performance decay of algorithms. Behind this, domain shift is one primary factors be blamed. In order tackle this problem, we propose using Potential Energy Ranking (PoER) decouple object feature given images, promoting label-discriminative representations while filtering out irrelevant correlations between objects background. PoER employs ranking loss shallow layers make features with identical category labels close each other vice versa. This makes neural networks aware both background characteristics, which vital for generating domain-invariant features. Subsequently, stacked convolutional blocks, further uses contrastive within categories distribute densely no matter domains, information progressively alignment. reports superior generalization benchmarks, improving average top-1 accuracy by at least 1.20% compared existing methods. Moreover, use ECCV 2022 NICO Challenge, achieving top place only vanilla ResNet-18 winning jury award. The code has been made publicly available at: https://github.com/ForeverPs/PoER.