A Comparative Study of Spatial Hierarchies and Pose-Aware Object Detection

• Author(s): Aditya Kinnori ; Aniket Tripathi
• Paper ID: 1709001
• Page: 186-197
• Published Date: 31-12-2020
• Published In: Iconic Research And Engineering Journals
• Publisher: IRE Journals
• e-ISSN: 2456-8880
• Volume/Issue: Volume 4 Issue 6 December-2020

Abstract

To understand the contents of a picture, computers in modern computer vision first focus on finding and spotting objects. If the environment is disorganized, people change positions, a part of their body is covered or other small details happen, CNNs find it tougher. Experts are now working with spatial hierarchy-based detection and pose-aware object detection as fresh techniques. The text examines both approaches, noting their key points, how they perform and the types of evaluation they fit for. They consider both the environment and the relationships around items to determine what they are by noticing how they appear and co-occur. They are most effective when there is enough supporting evidence for recognizing things. Another way to state it, pose-aware detection combines feature points and component detection, so non-rigid objects are detected better as their appearance depends on how they jointly move. Lately, Feature Pyramid Networks (FPN) are checked with graphs and Pose-RCNN alongside keypoint prediction are studied on the COCO 2017, PASCAL3D+, MPII and ADE20K datasets. Quantitative research shows that models with body-part awareness perform better than those with a fixed spatial structure and achieve up to 81.2% on keypoint tasks and higher mAP when dealing with dynamic poses. Even so, spatial models can process information fast and are precisely accurate with unchanging, multiple items. We also consider the speed at which a model makes a prediction (inference latency), how complicated the model is, any additional data it uses (annotations) and the extent to which it can be applied. It was determined that models that look at both location and pose can form a good base for future progress. Keep this document so you can learn and guide others about picking object-detection methods for various tasks.

Keywords

Spatial hierarchies, Pose-aware detection, Object detection, Deep learning, Convolutional neural networks, Feature representation, Computer vision, Visual perception, Contextual learning, Keypoint estimation.

Citations

IRE Journals:
Aditya Kinnori , Aniket Tripathi "A Comparative Study of Spatial Hierarchies and Pose-Aware Object Detection" Iconic Research And Engineering Journals Volume 4 Issue 6 2020 Page 186-197

IEEE:
Aditya Kinnori , Aniket Tripathi "A Comparative Study of Spatial Hierarchies and Pose-Aware Object Detection" Iconic Research And Engineering Journals, 4(6)