Inference Adaptive Thresholding based Non-Maximum Suppression for Object Detection in Video Image Sequence

TL;DR: Experimental results demonstrate that this simple and unsupervised method outperforms state-of-the-art NMS algorithms, with an increase of 6% in mean average precision (mAP) on the ImageNet VID dataset.

Abstract: This study proposes a novel inference adaptive thresholding based non-maximum suppression (NMS) (IAT-NMS) algorithm for deriving temporal cues between video sequences. The inference of temporal connectivity is first derived according to an overlapping measure of the bounding boxes between adjacent frames. Frames with high-confidence detection object are taken as key frames to leverage the scores of neighbor detections and preserve potential detections of blurred objects with low scores. Then, bounding boxes within each frame are ranked via their confidence scores and the overlapping ratio between the bounding box with the highest score against the remaining surrounding boxes is computed. This measure of overlapping is brought into a Gaussian function to estimate weights for adaptive suppression and to softly suppress the detection scores of possible severely overlapped objects. The proposed method is compared with state-of-the-art video object detection techniques. With the application of IAT-NMS, overlapping objects originally undistinguishable in the compared methods become detectable. Experimental results demonstrate that this simple and unsupervised method outperforms state-of-the-art NMS algorithms, with an increase of 6% in mean average precision (mAP) on the ImageNet VID dataset. Our study on performance limitations and sensitivity to parametric variations also finds that IAT-NMS demonstrates better detection capability than does the three compared algorithms, which fail to detect all targets or distinguish in the presence of multiple overlapping targets.