Researchers have developed a data-efficient method for segmenting crosswalks from overhead CCTV footage, addressing the challenge of viewpoint and appearance shifts compared to street-level imagery. The proposed pipeline utilizes a custom U-Net model trained on a limited set of annotated CCTV images and a larger pool of unlabeled frames. Pseudo-labeling techniques guided by confidence scores and geometric priors were employed to enhance performance, achieving an 88.91% IoU on manual validation data. The study also highlights the critical importance of isolating pseudo-label evaluation from the self-training data to ensure accurate performance metrics. AI
IMPACT This research offers a more efficient approach to training computer vision models for object detection in surveillance footage, potentially reducing annotation costs and improving real-world applicability.
RANK_REASON Academic paper detailing a novel methodology for computer vision task. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
