References & Citations
Computer Science > Computer Vision and Pattern Recognition
Title: Semi-Automatic Annotation For Visual Object Tracking
(Submitted on 18 Jan 2021 (v1), last revised 19 Aug 2021 (this version, v4))
Abstract: We propose a semi-automatic bounding box annotation method for visual object tracking by utilizing temporal information with a tracking-by-detection approach. For detection, we use an off-the-shelf object detector which is trained iteratively with the annotations generated by the proposed method, and we perform object detection on each frame independently. We employ Multiple Hypothesis Tracking (MHT) to exploit temporal information and to reduce the number of false-positives which makes it possible to use lower objectness thresholds for detection to increase recall. The tracklets formed by MHT are evaluated by human operators to enlarge the training set. This novel incremental learning approach helps to perform annotation iteratively. The experiments performed on AUTH Multidrone Dataset reveal that the annotation workload can be reduced up to 96% by the proposed approach.
Submission history
From: Aybora Koksal [view email][v1] Mon, 18 Jan 2021 10:35:32 GMT (1549kb,D)
[v2] Sun, 24 Jan 2021 13:20:59 GMT (1549kb,D)
[v3] Tue, 20 Apr 2021 09:49:20 GMT (1549kb,D)
[v4] Thu, 19 Aug 2021 20:29:09 GMT (2014kb,D)
Link back to: arXiv, form interface, contact.