Penalizing Proposals using Classifiers for Semi-Supervised Object Detection
This work addresses the cost and effort of obtaining annotated data for object detection by improving semi-supervised methods, though it is incremental as it builds on existing approaches with a specific modification.
The paper tackles the problem of poor performance in semi-supervised object detection due to machine-generated silver-standard labels by designing a modified loss function that incorporates a confidence metric for annotation quality. The result is a 4% gain in mAP with 25% labeled data and a 10% gain in mAP with 50% labeled data compared to a baseline without the confidence metric.
Obtaining gold standard annotated data for object detection is often costly, involving human-level effort. Semi-supervised object detection algorithms solve the problem with a small amount of gold-standard labels and a large unlabelled dataset used to generate silver-standard labels. But training on the silver standard labels does not produce good results, because they are machine-generated annotations. In this work, we design a modified loss function to train on large silver standard annotated sets generated by a weak annotator. We include a confidence metric associated with the annotation as an additional term in the loss function, signifying the quality of the annotation. We test the effectiveness of our approach on various test sets and use numerous variations to compare the results with some of the current approaches to object detection. In comparison with the baseline where no confidence metric is used, we achieved a 4% gain in mAP with 25% labeled data and 10% gain in mAP with 50% labeled data by using the proposed confidence metric.