In autonomous driving system, building a rigorous object detection model unaffected by conditions, such as weather or time-of-day, is essential for safety. However, as deep learning models are often limited in generalizability, training over the entire data collection can be suboptimal, e.g., daytime training instances hinder the training for nighttime prediction. We call this curse of multitasking (CoM), which was first observed in multilingual training, where training a multilingual model can be suboptimal, compared to multiple monolingual models. Our contribution is observing CoM in autonomous driving, overcoming the problem by building multiple mono-task models, or specialized experts for each task, then switching models according to the input condition, enhancing the overall effectiveness of the detection model. We show the effectiveness of using the proposed strategy in both YOLOv3 and RetinaNet models on BDD dataset.
|Number of pages||7|
|Journal||Journal of Computing Science and Engineering|
|Publication status||Published - 2022 Sept|
Bibliographical noteFunding Information:
This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2018-0-00769, Neuromorphic Computing Software Platform for Artificial Intelligence Systems).
© 2022. The Korean Institute of Information Scientists and Engineers
All Science Journal Classification (ASJC) codes
- Computer Science Applications