Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
102 tokens/sec
GPT-4o
59 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
50 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Scale Match for Tiny Person Detection (1912.10664v1)

Published 23 Dec 2019 in cs.CV

Abstract: Visual object detection has achieved unprecedented ad-vance with the rise of deep convolutional neural networks.However, detecting tiny objects (for example tiny per-sons less than 20 pixels) in large-scale images remainsnot well investigated. The extremely small objects raisea grand challenge about feature representation while themassive and complex backgrounds aggregate the risk offalse alarms. In this paper, we introduce a new benchmark,referred to as TinyPerson, opening up a promising directionfor tiny object detection in a long distance and with mas-sive backgrounds. We experimentally find that the scale mis-match between the dataset for network pre-training and thedataset for detector learning could deteriorate the featurerepresentation and the detectors. Accordingly, we proposea simple yet effective Scale Match approach to align theobject scales between the two datasets for favorable tiny-object representation. Experiments show the significantperformance gain of our proposed approach over state-of-the-art detectors, and the challenging aspects of TinyPersonrelated to real-world scenarios. The TinyPerson benchmarkand the code for our approach will be publicly available(https://github.com/ucas-vg/TinyBenchmark).(Attention: evaluation rules of AP have updated in benchmark after this paper accepted, So this paper use old rules. we will keep old rules of AP in benchmark, but we recommand the new and we will use the new in latter research.)

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (5)
  1. Xuehui Yu (23 papers)
  2. Yuqi Gong (10 papers)
  3. Nan Jiang (210 papers)
  4. Qixiang Ye (110 papers)
  5. Zhenjun Han (29 papers)
Citations (179)

Summary

Scale Match for Tiny Person Detection

The paper presents an innovative approach to addressing the challenges involved in detecting tiny objects, specifically tiny persons, in large-scale images, a task that remains underexplored in the field of computer vision. Detecting small objects is a critical aspect of many applications such as surveillance and search and rescue operations, where subjects may appear far away and amidst cluttered backgrounds.

Benchmark Overview: TinyPerson

The authors introduce "TinyPerson," a novel benchmark designed to focus on scenarios where individuals are at a substantial distance from the camera, resulting in subjects often appearing less than 20 pixels high. This benchmark emphasizes the difficulty posed by such reduced-size objects and the accompanying complex backgrounds that can lead to increased false detections. The TinyPerson dataset sets a precedent as the first benchmark aiming at facilitating detection in long-distance settings with extensive backgrounds, particularly in maritime contexts.

Key Contributions

The paper's primary contributions include:

  1. TinyPerson Benchmark Introduction: TinyPerson is positioned as a significant dataset, addressing the understudied area of tiny-person detection. Annotations are provided for over 72,000 objects, segmented into various subcategories and considerations such as sea and land-based persons.
  2. Scale Match Approach: A substantial solution proposed is the "Scale Match" technique. The methodology involves aligning the scale distribution of pre-training and detector-training datasets to enhance feature representation consistency, addressing the scale mismatch issue that undermines detection performance.
  3. Improved Detection Results: The authors demonstrate a significant performance boost (~5%) over state-of-the-art detectors using their proposed Scale Match on the challenging TinyPerson dataset. Their results underscore the difficulty of achieving precise detection with tiny objects due to both absolute and relative size challenges.

Experimental Insights

The paper provides a detailed analysis of the dataset's challenges, such as tiny absolute and relative sizes, which aggravate detection difficulties compared to other datasets like COCO or CityPersons. Standard models like Faster RCNN-FPN, FCOS, and RetinaNet were evaluated, exhibiting varied performance decreases when confronted with tiny-scale person detection scenarios. Their analysis emphasizes the necessity of preserving spatial information and effectively managing dense anchor points.

Implications and Future Directions

The presented Scale Match algorithm optimizes size alignment, effectively utilizing additional data across different scales, which holds potential beyond the scope of tiny object detection, extending to diverse object detection tasks with varying size distributions. This scale alignment strategy may influence future research in areas such as transfer learning and domain adaptation, where aligning feature representations is critical.

Potential future developments include enhancing the Scale Match method to incorporate additional contextual understanding beyond scale, further refining object segmentation and classification tasks amid cluttered backgrounds and exploring broader applications in different object detection fields.

In conclusion, the paper offers a valuable methodological advancement with the introduction of the TinyPerson benchmark and the innovative Scale Match approach. These contributions provide critical insights and tools for researchers tackling the complex challenges inherent in tiny object detection, paving the way for improved systems capable of handling real-world environments where object sizes vary significantly.

Github Logo Streamline Icon: https://streamlinehq.com