C-SFDA: A Curriculum Learning Aided Self-Training Famework for Efficient Source Free Domain Adaptation
Nazmul Karim, Niluthpol Chowdhury Mithun, Abhinav Rajvanshi, Han-pang Chiu, Supun Samarasekera,
Nazanin Rahnavard
Department of ECE, UCF, Orlando, FL, USA SRI International, Princeton, NJ, USA
Accepted at CVPR 2023
Abstract
Unsupervised domain adaptation (UDA) approaches focus on adapting models trained on a labeled source domain to an unlabeled target domain. In contrast to UDA, source-free domain adaptation (SFDA) is a more practical setup as access to source data is no longer required during adaptation. Recent state-of-the-art (SOTA) methods on SFDA mostly focus on pseudo-label refinement based self-training which generally suffers from two issues: i) inevitable occurrence of noisy pseudo-labels that could lead to early training time memorization, ii) refinement process requires maintaining a memory bank which creates a significant burden in resource constraint scenarios. To address these concerns, we propose C-SFDA, a curriculum learning aided self-training framework for SFDA that adapts efficiently and reliably to changes across domains based on selective pseudo-labeling. Specifically, we employ a curriculum learning scheme to promote learning from a restricted amount of pseudo labels selected based on their reliabilities. This simple yet effective step successfully prevents label noise propagation during different stages of adaptation and eliminates the need for costly memory-bank based label refinement. Our extensive experimental evaluations on both image recognition and semantic segmentation tasks confirm the effectiveness of our method. C-SFDA is readily applicable to online test-time domain adaptation and also outperforms previous SOTA methods in this task.
Overview
Figure 1. Left: In source-free domain adaptation, we only have a source model that needs to be adapted on the target data. Among the source-generated pseudo-labels, a large portion is noisy which is important to avoid during supervised self-training (SST) with regular cross-entropy loss. Instead of using all pseudo-labels, we choose the most reliable ones and effectively propagate high-quality label information to unreliable samples. As the training progresses, the proposed selection strategy tends to choose more samples for SST due to the improved average reliability of pseudo-labels. Such a restricted self-training strategy creates a model with better discrimination ability and eventually corrects the noisy predictions. Here, T is the total number of iterations. Right: While existing SFDA techniques leverages cluster structure knowledge in the feature space, there may exist many misleading neighbors–neighbors’ pseudo labels that are different from the anchors’ true label. Therefore, clustering-based label propagation inevitably suffers from label noise in subsequent training.
Summary of Contributions
We introduce a novel SFDA technique that focuses on noise-free self-training exploiting the reliability of generated pseudo-labels. With the help of curriculum learning, we aim to prevent early training time memorization of noisy pseudo-labels and improve the quality of subsequent self-training as shown in Figure 1.
By prioritizing the learning from highly reliable pseudo-labels first, we aim to propagate refined and accurate label information among less reliable samples. Such a selective self-training strategy eliminates the requirement of a computationally costly and memory-bank dependent label refinement framework.
C-SFDA achieves state-of-the-art performance on major benchmarks for image recognition and semantic segmentation. Being highly memory-efficient, the proposed method is readily applicable to online test-time adaptation settings and obtains SOTA performance.
Proposed Method
After calculating the pseudo-labels, we perform selective pseudo-labeling to separate them into reliable and unreliable groups.
We apply three different losses to learn in a curriculum manner by prioritizing learning from the reliable group of samples first.
We apply L number strong augmentations which helps both in pseudo-label refinement and contrastive learning.
Performance
VISDA-C Dataset