Density-Guided Dense Pseudo Label Selection For Semi-supervised Oriented Object Detection

TL;DR

In DDPLS, a simple but effective adaptive mechanism to guide the selection of dense pseudo labels is designed and the Pseudo Density Score (PDS) is proposed to estimate the density of potential objects and use this score to select reliable dense pseudo labels.

Abstract

Recently, dense pseudo-label, which directly selects pseudo labels from the original output of the teacher model without any complicated post-processing steps, has received considerable attention in semi-supervised object detection (SSOD). However, for the multi-oriented and dense objects that are common in aerial scenes, existing dense pseudo-label selection methods are inefficient because they ignore the significant density difference. Therefore, we propose Density-Guided Dense Pseudo Label Selection (DDPLS) for semi-supervised oriented object detection. In DDPLS, we design a simple but effective adaptive mechanism to guide the selection of dense pseudo labels. Specifically, we propose the Pseudo Density Score (PDS) to estimate the density of potential objects and use this score to select reliable dense pseudo labels. On the DOTA-v1.5 benchmark, the proposed method outperforms previous methods especially when labeled data are scarce. For example, it achieves 49.78 mAP given only 5\% of annotated data, which surpasses previous state-of-the-art method given 10\% of annotated data by 1.15 mAP. Our codes is available at https://github.com/Haru-zt/DDPLS.

Figures (5)

• Figure 1: Significant density difference in DOTA-v1.5 compared with COCO
• Figure 2: The overview of proposed DDPLS. Each training batch consists of both labeled data and unlabeled data. Note that we hide the supervised part for simplicity. For the unsupervised part, we sample dense pseudo labels according to the Density-Guided Dense Pseudo Label Selection.
• Figure 3: The correlation between the Pseudo Density Score and the relative number of pseudo labels selected under the 10% setting. Relative number indicates the sum of confidence of pseudo labels selected.
• Figure 4: We randomly sampled 1k unlabeled training images under the 10% setting and calculated the average Pseudo Density Score under different train iterations.
• Figure 5: Some visualization examples from DOTA-v1.5 dataset. The green rectangles indicate predictions. The red dashed circle, solid red circle, and red arrow represent false negative, false positive, and inaccurate orientation prediction, respectively.