Uncertainty Aware Training to Improve Uncertainty Active Learning for Semantic Segmentation

Moritz Thoma, Tobias Preintner, Emad Aghajanzadeh, Shambhavi Balamuthu Sampath, Pierpaolo Mori, Nael Fasfous, Manoj-Rohit Vemparala, Alexander Frickenstein, Daniel Mueller-Gritschneder, Ulf Schlichtmann; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025, pp. 4410-4420

Active learning (AL) on deployed devices involves collecting data to improve the performance of machine learning models beyond their initial version. In this context, the collected data must be uploaded to a remote server for training the next version of the model. Data collection on deployed devices can pose one of two challenges: (1) if done naively, large quantities of uninformative data are uploaded or (2) if complex data selection algorithms are used, careful curation of data introduces a compute overhead that is unrelated to the main task. In this paper, we introduce a novel Uncertainty Aware Active Learning (UAAL) method for semantic segmentation that integrates uncertainty estimates into the training process, thereby conditioning the model for an AL setup before deployment time. This enhances the quality of the collected data for an existing selection technique without adding more complexity to the selection algorithm. Essentially, the model outputs provide a clearer indication to the selection algorithm about its own uncertainty. We integrate UAAL with five uncertainty AL methods and demonstrate its efficacy on the CityScapes and ADE20K datasets using three DeepLabv3+ variants. On the CityScapes dataset, with a 10 % data budget, UAAL achieves an average mIoU increase of +1.74 p.p., peaking at +3.49 p.p. for MobileNetV3 combined with MC-Dropout. Similarly, on ADE20K, UAAL boosts ResNet-50 with MC-Dropout by +2.22 p.p. at the same data budget, all while not adding any complexity to the data selection process.