Domain Adaptive Semantic Segmentation without Source Data

Abstract

Domain adaptive semantic segmentation is recognized as a promising technique to alleviate the domain shift between the labeled source domain and the unlabeled target domain in many real-world applications, such as automatic pilot. However, large amounts of source domain data often introduce significant costs in storage and training, and sometimes the source data is inaccessible due to privacy policies. To address these problems, we investigate domain adaptive semantic segmentation without source data, which assumes that the model is pre-trained on the source domain, and then adapting to the target domain without accessing source data anymore. Since there is no supervision from the source domain data, many self-training methods tend to fall into the "winner-takesall" dilemma, where the majority classes totally dominate the segmentation networks and the networks fail to classify the minority classes. Consequently, we propose an effective framework for this challenging problem with two components: positive learning and negative learning. In positive learning, we select the class-balanced pseudo-labeled pixels with intra-class threshold, while in negative learning, for each pixel, we investigate which category the pixel does not belong to with the proposed heuristic complementary label selection. Notably, our framework can be easily implemented and incorporated with other methods to further enhance the performance. Extensive experiments on two widely-used synthetic-to-real benchmarks demonstrate our claims and the effectiveness of our framework, which outperforms the baseline with a large margin. Code is available at https://github.com/fumyou13/LDBE.

Iterative Loop Method Combining Active and Semi-Supervised Learning for Domain Adaptive Semantic Segmentation

“…We further evaluate the generalization of STAL by extending to source-free domain adaptation (SFDA). Comparing methods include URMA [19], LD [74], SFDA [36], and [58] 64.1 -CCT (3.1%) [48] 66.4 -Cutmix (3.1%) [20] 69.1 -AEL (3.1%) [32] 74.3 -U2PL+AEL (6.3%) [9] 74.9 - t-SNE Visualization. To better develop intuition, we draw t-SNE visualization [61] of the learned feature representations for contrast methods (RIPU [70]) and ours STAL in Fig.…”

Source-Free Domain Adaptation for Real-World Image Dehazing

“…Qu et al [91] use the idea of multiple instance learning [96], [97] to form a balanced global sampling strategy. You et al [98] set category-specific thresholds to keep the number between categories as consistent as possible, Li et al [99] noticed that the model may be biased towards most categories, thus proposing an imbalanced SFDA strategy with secondary label correction. Pseudo-label Filtering.…”

Cross-Modality Domain Adaptation for Freespace Detection: A Simple yet Effective Baseline

“…Since directly evaluating on one domain with a model trained on another domain usually causes performance drops, domain adaptation methods are developed to help address the domain shift between two distinct domains. Domain adaptation is widely used in classification [14,17,21,64], object detection [8,33,52,65,66], semantic segmentation [20,27,57,71,78] and many other fields [39,72,74,75]. In particular, unsupervised domain adaptation (UDA) methods have attracted substantial attention since they are free from manuallyannotated labels in the target domain.…”