Bag of Tricks for “Vision Meet Alage” Object Detection Challenge

Abstract

In this paper, we introduce our solution to the “Vision Meets Algae” Workshop and Challenge (VisAlgae) in details. Since a large number of small objects and similar categories, the location and classification of algae are challenging. For that, we propose a bag of tricks for VisAlgae, including data augmentation, model architecture, and pipeline. For data augmentation, we introduce bounding-box jitter, mix-up, multi-scale, albu, and test time augmentation to increase sample diversity and randomness. We learn a better region of interest (RoI) features by adding global semantic information to RoI features. Especially a novelty double head is devised to enhance final features via reserving spatial and channel information. For the pipeline, We introduce the detector framework, backbone, stochastic weights averaging, pseudo labels, and weighted boxes fusion. Experimental results demonstrate that our approach can achieve an excellent mean average precision (mAP) performance of object detection.