| Presentation | 2022-12-05 Accuracy Improvement of Small Object Detection Based on Deep Learning Junya Morioka, Ryusuke Miyamoto, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Various methods based on deep learning have been proposed for object detection, but there is still much room for improving the accuracy of small object detection. In this paper, we revisit the fundamentals of visual object detection based on deep learning and discuss how to improve its accuracy. Comparison of the Bird, SAVMAP, and VisDrone dataset dedicated to small object detection with the generic PascalVOC and COCO datasets indicates that the IoU corresponding to intersection of a ground truth and a detection result is very low for small object detection. Moreover, experimental results using RetinaNet, EfficientDet, and YOLOv5 showed that the input size, anchor box, depth of model layers, and surrounding context of target objects were important factors for small object detection. In addition, it is also shown that the accuracy of small object detection depends on the anchor box more than generic object detection, shallow layer close to the detector input contains significant information for small object detection, and the significant information maybe lost as the layers of the model layers become deeper. Experimental results showed that detection accuracy was improved by the following options: enlarging bounding boxes, using higher resolution input, and network architecture with shallower layers. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | Object Detection / Small Object Detection / Deep Learning |
| Paper # | SIS2022-29 |
| Date of Issue | 2022-11-28 (SIS) |
| Conference Information | |
| Committee | SIS |
|---|---|
| Conference Date | 2022/12/5(2days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | |
| Chair | Tomoaki Kimura(Kanagawa Inst. of Tech.) |
| Vice Chair | Naoto Sasaoka(Tottori Univ.) / Hakaru Tamukoh(Kyushu Inst. of Tech.) |
| Secretary | Naoto Sasaoka(NTT) / Hakaru Tamukoh(Kansai Univ.) |
| Assistant | Yoshiaki Makabe(Kanagawa Inst. of Tech.) / Yosuke Sugiura(Saitama Univ.) |
| Paper Information | |
| Registration To | Technical Committee on Smart Info-Media Systems |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Accuracy Improvement of Small Object Detection Based on Deep Learning |
| Sub Title (in English) | |
| Keyword(1) | Object Detection |
| Keyword(2) | Small Object Detection |
| Keyword(3) | Deep Learning |
| 1st Author's Name | Junya Morioka |
| 1st Author's Affiliation | Meiji University(Meiji Univ.) |
| 2nd Author's Name | Ryusuke Miyamoto |
| 2nd Author's Affiliation | Meiji University(Meiji Univ.) |
| Date | 2022-12-05 |
| Paper # | SIS2022-29 |
| Volume (vol) | vol.122 |
| Number (no) | SIS-293 |
| Page | pp.pp.32-37(SIS), |
| #Pages | 6 |
| Date of Issue | 2022-11-28 (SIS) |