Object detection in car cabin environment
Attention mechanisms and bounding box regression (BBR) losses have been widely used for object detection in the car cabin environment, achieving remarkable improvements in feature extraction and prediction. However, most existing research has not systematically studied these two components, neglecti...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis-Master by Coursework |
| Language: | English |
| Published: |
Nanyang Technological University
2024
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/174052 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Nanyang Technological University |
| Language: | English |
| Summary: | Attention mechanisms and bounding box regression (BBR) losses have been widely used for object detection in the car cabin environment, achieving remarkable improvements in feature extraction and prediction. However, most existing research has not systematically studied these two components, neglecting to explore their potential interactions. To mitigate the adverse effects caused thereby, we have not only devised a novel attention mechanism and a unique BBR loss but also demonstrated their synergistic effect. Firstly, a Deformable Coordinate Attention (DCA) is proposed, leveraging deformable convolution to extract features more flexibly in both channel and spatial dimensions. Secondly, a Step Efficient Intersection over Union (SEIOU) loss is designed to achieve high-efficiency BBR. Finally, extensive experimentations on the Drive and Act, MS COCO detection, PASCAL VOC 2007 detection, and
PASCAL VOC 2012 detection dataset reveal the synergistic effect between DCA and SEIOU in object detection tasks. Notably, our modules can be flexibly plugged into classical networks with minimal computational overhead. |
|---|