Asymmetric Loss For Multi-Label Classification

Abstract

Pictures of everyday life are inherently multi-label in nature. Hence, multi-label classification is commonly used to analyze their content. In typical multi-label datasets, each picture contains only a few positive labels, and many negative ones. This positive-negative imbalance can result in under-emphasizing gradients from positive labels during training, leading to poor accuracy.In this paper, we introduce a novel asymmetric loss ("ASL"), that operates differently on positive and negative samples. The loss dynamically down-weights the importance of easy negative samples, causing the optimization process to focus more on the positive samples, and also enables to discard mislabeled negative samples.We demonstrate how ASL leads to a more "balanced" network, with increased average probabilities for positive samples, and show how this balanced network is translated to better mAP scores, compared to commonly used losses. Furthermore, we offer a method that can dynamically adjust the level of asymmetry throughout the training.With ASL, we reach new state-of-the-art results on three common multi-label datasets, including achieving 86.6% on MS-COCO. We also demonstrate ASL applicability for other tasks such as fine-grain single-label classification and object detection.

Simple and Robust Loss Design for Multi-Label Learning with Missing Labels

“…The reason behind this phenomenon is the distinctive features in multi-label learning: 1) the domination of negatives on positives in multiple binary classifiers; In the multi-label setting, an image typically contains a few positives but much more negatives, which results in serious positive-negative imbalance [26]. 2) Missing labels exacerbate the positivenegative imbalance and plague the learning of recognizing positives.…”

“…VOC [31] contains 5,717 training images with 20 classes, and additional 5,823 images are used for test. Following [26], we collect NUS-wide [32] with 119,103 and 50,720 images as training set and test set, consists of 81 classes. Due to many download urls of Openimages [33] are expired, we were able We construct the training sets of missing labels by randomly dropping positive labels for each training image with different ratios.…”

“…2) Evaluation metrics: Following previous multi-label image classification researches [6], [26], mean Average Precision (mAP) is used as the primary metric to evaluate model performance. We also report the overall F1-measure (OF1) and per-category F1-measure (CF1) with the positive threshold of 0.5.…”

“…The optimizer is the Adam with a weight decay of 1e-4 and cycle learning rate schedule [35] is used with the max learning rate of 1e-4. Following [26], we use the exponential moving average trick for higher baselines.…”

“…Weak Assume Negatives (WAN) and Label Smoothing (LS) [6] are selected as the stronger baselines than BCE in MLML. Secondly, we compare the Hill loss with loss re-weighting methods, Focal [15] and the recently proposed ASL [26]. Thirdly, we choose the loss correction methods, including Pseudo Label [23] and Regularized Online Label Estimation (ROLE) [6] to validate the effectiveness of the SPLC.…”