The 1st Tiny Object Detection Challenge:Methods and Results

Abstract: The 1st Tiny Object Detection (TOD) Challenge aims to encourage research in developing novel and accurate methods for tiny object detection in images which have wide views, with a current focus on tiny person detection. The TinyPerson dataset was used for the TOD Challenge and is publicly released. It has 1610 images and 72651 box-levelannotations. Around 36 participating teams from the globe competed inthe 1st TOD Challenge. In this paper, we provide a brief summary of the1st TOD Challenge including brief introductions to the top three methods.The submission leaderboard will be reopened for researchers that areinterested in the TOD challenge. The benchmark dataset and other information can be found at: https://github.com/ucas-vg/TinyBenchmark.

• The paper introduces the first Tiny Object Detection Challenge, highlighting the innovative use of the TinyPerson dataset and detailed evaluation metrics.
• It presents methodologies leveraging multi-scale data augmentation, two-stage frameworks, and ensemble strategies to enhance detection performance.
• The results underscore the challenges of tiny object detection and suggest future research directions for improving detection in diverse and complex scenarios.

Overview of the 1st Tiny Object Detection Challenge: Methods and Results

The work by Xuehui Yu et al. presents a comprehensive account of the inaugural Tiny Object Detection (TOD) Challenge. The challenge was designed to stimulate research in the nascent but crucial area of tiny object detection, using the TinyPerson dataset which focuses on identifying small-sized human figures in images with large field-of-view.

Challenge Dataset and Evaluation

The TinyPerson dataset, comprising 1,610 images and 72,651 annotated instances, served as the central element of this challenge. These images were extracted from real-world video footage and included distinct categories such as “sea person” and “earth person” based on their spatial context within the image. The dataset’s classifications also included "ignore" regions where highly dense or ambiguous annotations made traditional bounding box approaches infeasible.

The evaluation metrics utilized were Average Precision (AP) and Miss Rate (MR). The tiny object size interval was further subdivided to enable focused analysis of detection capabilities across varying object scales, with particular emphasis on tiny1, tiny2, and tiny3 sub-categories based on pixel area.

Challenge Results and Top Performers

Thirty-six teams participated in the challenge, with outcomes detailed in Table 1 ranking teams by their $AP_{50}^{tiny}$ scores. Notably, the top-performing teams employed sophisticated methodologies that integrated various detection frameworks, data augmentation techniques, and advanced ensemble strategies.

Key Insights from the Top Teams

1. Baidu_ppdet
   • Utilized a two-stage detection framework with a diverse model ensemble.
   • Enhanced model training with large-scale data and innovative feature fusion techniques.
   • Achieved superior object detection by leveraging scale matchmaking and strategic model refinements.
2. STY-402
   • Adopted Faster R-CNN architectures with enriched backbone networks to enhance detection robustness.
   • Implemented comprehensive data augmentation and multi-scale training strategies, bolstered by additional dataset usage.
3. BRiLliant
   • Focused on high-resolution feature representation using HRNet.
   • Introduced an improved CBAM mechanism to assist in precise region proposal generation.
   • Developed an adaptive sampling strategy to adjust to the variety in object scales, optimizing small object detection.

Implications and Future Research Directions

The results from the TOD Challenge underline the inherent complexity of tiny object detection, a domain crucial for applications such as surveillance, autonomous navigation, and aerial imagery analysis. The methods explored by the top-performing teams demonstrate potential pathways to overcoming the unique challenges posed by small-scale object detection.

Going forward, expanding this research to encompass a broader variety of tiny object categories beyond human figures will be essential. Moreover, integrating multi-modal data and exploring novel deep learning architectures could significantly enhance detection capabilities. The challenge highlights that advancements in this field could also drive progress in general object detection research, as many strategies developed for tiny object scenarios are adaptable to broader contexts.

The TOD Challenge marks an important step in recognizing and elevating the study of tiny object detection within the computer vision community, setting precedents for future contests and research initiatives.