循序漸進：用分階段訓練提升半監督學習方法在領域自適應姿態估計中的適用性

Abstract

近幾年，領域自適應姿態估計 (DAPE) 受到越來越多的關注。在解決該任務的方法中，半監督式學習 (SSL) 的方法因為與無監督領域自適應 (UDA) 有相似的目標而被廣泛使用。然而，雖然 SSL 和 UDA 同樣都旨在用未標記的數據增強已在標記數據上訓練的模型，兩個任務對於資料分布的不同期望依然讓 SSL 方法無法完美地契合於 UDA 中。有鑑於此，我們提出了 Phase2Phase，這一策略整合了三種方法：Adaptive Mean Teacher、T-VAT-based UniMatch，以及 Mixup Augmentation。 傳統的 Mean Teacher 方法通常採用較大的平滑係數，但在 UDA 任務中，較大的平滑細數容易阻礙教師模型迅速達到學生模型的性能。為了克服這一問題，我們提出 Adaptive Mean Teacher。藉由在領域自適應初期加入緩衝階段，並在該緩衝階段使用較低的平滑係數，我們使教師模型在保持 Mean Teacher 穩定性特點的同時，使其在初期就能快速追上學生模型的表現。 Mean Teacher 已經被證實在 UDA 的情境下有效，而除了如 Mean Teacher 這樣以模型為中心出發，用不同模型生成兩個輸出以進行一致性調節的方法，SSL 中同樣存在像 FixMatch 和 UniMatch 這樣在輸入和特徵層面引入干擾項的方法。儘管這些策略理論上能相輔相成，但實驗卻顯示生硬地結合兩個方法會導致效能下降。為此，我們提出的 Phase2Phase 在採用傳統 Mean Teacher 的訓練階段後又引入了一個額外階段，在該階段利用 T-VAT-based UniMatch 讓模型校能進一步提升。最後，我們還在訓練全程都引入 Mixup 技術，以提升模型的穩健性和整體性能。 實驗結果顯示，我們提出的方法確實有益於提升 DAPE 的表現。值得注意的是，Phase2Phase 超越了能使用標記資料的其他 DAPE 方法，同時他的表現也逼近當前最先進、在不使用標記資料的情況下依然表現優異 DAPE 方法，SFDAHPE。這凸顯了 Phase2Phase 在實際情境中的實用性。

Domain Adaptive Pose Estimation (DAPE) has received increasing attention recently. Many methods commonly used in semi-supervised learning (SSL) tasks are now applied to DAPE tasks because SSL shares a similar goal with unsupervised domain adaptation (UDA). Although both SSL and UDA aim to enhance a model trained on labeled data with unlabeled data, the differences in data distribution expectations present a challenge, preventing these SSL methods from seamlessly adapting to DAPE. With an awareness of inconsistent expectations in the data, we introduce Phase2Phase, a novel approach that integrates three core strategies: Adaptive Mean Teacher, T-VAT-based UniMatch, and Mixup augmentation. In traditional Mean Teacher methods, a significant smoothing coefficient is typically employed. However, in UDA tasks, a large smoothing coefficient can hinder the teacher model from quickly achieving the performance of the student model. To overcome this issue, we propose the Adaptive Mean Teacher. By introducing a ramp-up phase during the initial stages of domain adaptation, where a reduced smoothing coefficient is applied, we enable the teacher model to rapidly align with the performance of the student model, while preserving the stability inherent in the traditional Mean Teacher framework. Mean Teacher has been proven effective in UDA, and in addition to model-centered approaches like Mean Teacher, which generate two outputs from different models for consistency regularization, SSL also includes methods like FixMatch and UniMatch that introduce disturbances at the input and feature levels. Although these strategies theoretically complement each other, their direct combination can lead to degraded effectiveness. Our thesis introduces an additional phase for T-VAT-based UniMatch to facilitate integration. Finally, we incorporate Mixup augmentation to boost the model's robustness, further elevating the overall performance. The experimental results show that all the proposed methods enhance the performance of DAPE. Notably, Phase2Phase surpasses previous source-dependent DAPE approaches and achieves comparable results to the current state of the art in source-free DAPE, SFDAHPE. This underscores Phase2Phase’s practical effectiveness in the real-world scenarios.