基於圖像級標籤的類增量持續學習運用於實例分割

Abstract

以往的實例分割任務需要對圖像中復雜的物體輪廓進行複雜的標註以進行訓練。實際上，標籤也能以逐步的方式提供，以在不同時間步驟中平衡所需的人力資源。然而，目前在實例分割任務上，僅使用弱標籤的增量學習研究仍較沒受到討論。因此在本文中，我們提出了一個持續學習方法，利用圖像級別標籤來分割出物體實例。與大多數依賴傳統方法提供物件候選框的弱監督實例分割不同，我們將弱監督語義分割中的語義知識轉移到弱監督實例分割中生成實例線索。為了解決持續學習中的背景偏移問題，我們利用由過去模型生成的舊類別分割結果來提供更可靠的語義、實例和峰值假設。據我們所知，這是首篇使用圖像標籤的弱監督持續學習實例分割的工作。實驗結果顯示，在不同的增量設置下，我們的方法在Pascal VOC和COCO資料集上都可以達到更好的表現。

Instance segmentation in computer vision requires time-consuming manual labeling of complex object contours in images for training. To alleviate the labor-intensive nature of this process, incremental labeling techniques have been explored to distribute human effort across different time steps. However, the field of incremental learning for instance segmentation with weak labels remains underdeveloped. In this study, we propose a novel method for continual learning that enables object instance segmentation using only image-level labels. In contrast to existing weakly-supervised instance segmentation (WSIS) methods that heavily rely on traditional object proposals, we leverage the knowledge gained from weakly-supervised semantic segmentation (WSSS) to generate valuable instance cues. To address the challenge of background shift in continual learning for segmentation, we incorporate the results of previously trained models for old classes. This integration provides more reliable semantic, instance and peak hypotheses during the learning process. To the best of our knowledge, this work is the first to explore weakly-supervised continual learning for instance segmentation in images. Through extensive experiments conducted on the Pascal VOC and COCO datasets, we demonstrate that our proposed method can achieve better performance under various incremental learning settings.