Name
Title | ||
|---|---|---|
Utilizing incremental branches on a one-stage object detection framework to avoid catastrophic forgetting |
Abstract | ||
|---|---|---|
The tremendous success of deep learning on object detection tasks compels researchers to adopt deep learning models for autonomous driving vehicles. As autonomous driving vehicles grows sophisticated, the top models are expected to detect novel classes beyond its prior objectives. Thus, incremental learning for object detection essentially ensures that a model is able to detect additional classes on the fly. In this work, we demonstrate how to update a model on new data and an existing model to append new classes on the existing model. The proposed method utilizes episodic memory to save finite samples of data and replay them during incremental learning. The results on PASCAL VOC2007 have suggested that the proposed method obtains the least mAP reduction, at 4.3%, compared against the all-classes learning in the 10+10 classes scenario, which is the lowest amongst other prior arts. Our method also has the highest backward and forward transfer among incremental learning strategies, indicating better memorization and adaptability. |
Year | DOI | Venue |
|---|---|---|
2022 | 10.1007/s00138-022-01284-z | MACHINE VISION AND APPLICATIONS |
Keywords | DocType | Volume |
Incremental learning, One-stage object detection, Imbalanced data, Autonomous driving assistance systems | Journal | 33 |
Issue | ISSN | Citations |
2 | 0932-8092 | 0 |
PageRank | References | Authors |
0.34 | 0 | 5 |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Jeng-Lun Shieh | 1 | 0 | 0.34 |
| Muhamad Amirul Haq | 2 | 0 | 0.34 |
| Qazi Mazhar ul Haq | 3 | 0 | 0.34 |
| Shanq-Jang Ruan | 4 | 375 | 55.44 |
| Peter Chondro | 5 | 7 | 4.27 |