Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
125 tokens/sec
GPT-4o
53 tokens/sec
Gemini 2.5 Pro Pro
42 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Towards Large-Scale Small Object Detection: Survey and Benchmarks (2207.14096v4)

Published 28 Jul 2022 in cs.CV

Abstract: With the rise of deep convolutional neural networks, object detection has achieved prominent advances in past years. However, such prosperity could not camouflage the unsatisfactory situation of Small Object Detection (SOD), one of the notoriously challenging tasks in computer vision, owing to the poor visual appearance and noisy representation caused by the intrinsic structure of small targets. In addition, large-scale dataset for benchmarking small object detection methods remains a bottleneck. In this paper, we first conduct a thorough review of small object detection. Then, to catalyze the development of SOD, we construct two large-scale Small Object Detection dAtasets (SODA), SODA-D and SODA-A, which focus on the Driving and Aerial scenarios respectively. SODA-D includes 24828 high-quality traffic images and 278433 instances of nine categories. For SODA-A, we harvest 2513 high resolution aerial images and annotate 872069 instances over nine classes. The proposed datasets, as we know, are the first-ever attempt to large-scale benchmarks with a vast collection of exhaustively annotated instances tailored for multi-category SOD. Finally, we evaluate the performance of mainstream methods on SODA. We expect the released benchmarks could facilitate the development of SOD and spawn more breakthroughs in this field. Datasets and codes are available at: \url{https://shaunyuan22.github.io/SODA}.

Citations (218)

Summary

  • The paper introduces two large-scale datasets (SODA-D and SODA-A) to address data scarcity and benchmark small object detection methods.
  • It categorizes existing detection techniques into six groups, including sample-oriented, scale-aware, and attention-based methods.
  • The experimental analysis using AP metrics highlights performance gaps in detecting small objects and outlines future research directions.

This paper presents a comprehensive survey and benchmark for the task of Small Object Detection (SOD). The authors highlight the challenges associated with SOD, including poor visual appearance, noisy representations, and the lack of large-scale datasets for training and evaluation. To address the data scarcity issue, they introduce two new large-scale datasets, SODA-D and SODA-A, specifically designed for SOD in driving and aerial scenarios, respectively. The paper thoroughly reviews existing SOD methods, categorizing them into six major groups, and evaluates the performance of mainstream object detection algorithms on the newly created SODA datasets.

Here's a breakdown of the key aspects:

1. Introduction and Motivation:

  • The paper emphasizes the importance of SOD in various applications, such as surveillance, drone scene analysis, and autonomous driving.
  • It points out the significant performance gap between detecting small and normal-sized objects, even with state-of-the-art detectors.
  • The authors attribute this gap to the inherent difficulties in learning representations from limited and distorted information of small objects, as well as the lack of suitable datasets.
  • The paper proposes to create large-scale datasets with exhaustively annotated small objects to facilitate the development and evaluation of SOD algorithms.

2. Challenges of Small Object Detection:

  • Information Loss: Down-sampling operations in CNNs diminish the spatial redundancy and high dimensional features, but also extinguishes the representation of tiny objects.
  • Noisy Feature Representation: Small objects have low resolution and poor-quality appearance, making it difficult to learn discriminative features. Features are often contaminated by background noise.
  • Low Tolerance for Bounding Box Perturbation: Slight deviations in bounding box predictions significantly impact the Intersection over Union (IoU) for small objects.
  • Inadequate Samples for Training: Small object regions have limited overlap to priors (anchors or points), leading to insufficient positive samples.

3. Review of Small Object Detection Algorithms:

The paper categorizes existing deep learning-based SOD methods into six groups:

  • Sample-oriented Methods: Focus on increasing the number of small object samples through data augmentation or optimizing the assignment strategy to ensure sufficient samples for network training. (e.g., AdaResampling, methods using DS-GAN, S3FD, RFLA)
  • Scale-aware Methods: Aim to construct scale-specific detectors or fuse hierarchical features to improve representation. (e.g., MS-CNN, FPN, SNIP, Sniper, IPG-Net)
  • Attention-based Methods: Leverage visual attention mechanisms to highlight important regions and suppress unnecessary ones. (e.g., SCRDet, FBR-Net, KB-RANN, CANet)
  • Feature-imitation Methods: Enrich the features of small objects by mimicking those of larger objects, using similarity learning or super-resolution techniques. (e.g., Self-Mimic Learning, LPR Memory, MTGAN, EE-GAN, SRC-GAN, PerceptualGAN)
  • Context-modeling Methods: Exploit contextual cues (semantic or spatial associations) to improve the detection of small objects. (e.g., methods using contextual regions around proposal patches, PyramidBox, FS-SSD, SINet, CAD-Net)
  • Focus-and-detect Methods: Filter out regions without objects to reduce computation and focus on relevant areas. (e.g., ClusDet, EdgeDuet, FS)

4. Review of Datasets for Small Object Detection:

  • The paper reviews several publicly available datasets that contain a considerable number of small objects, including COCO, WiderFace, TinyPerson, TT100K, VisDrone and DOTA.
  • It points out that existing datasets are either designed for single-category detection tasks or have small objects distributed in only a few categories.
  • The authors argue that these datasets do not adequately support the training of deep learning models specifically customized for multi-category SOD.

5. SODA Datasets (SODA-D and SODA-A):

  • Data Acquisition and Annotation: Details how the datasets were created including the definition of a "valuable object" based on pixel area, data sources, dataset splits (train/val/test), category selection, and annotation process using tools like LabelImg and LabelMe. Defines "small" objects as <= 1024 pixels.
  • Statistical Analysis: Provides statistics on the number of images, instances per category, and the distribution of object sizes.
  • SODA-D: Focuses on driving scenarios and includes 24,828 images with 278,433 instances of nine categories (people, rider, bicycle, motor, vehicle, traffic-sign, traffic-light, traffic-camera, and warning-cone). Key features include rich diversity, high spatial resolution, and a large number of "ignore" regions to exclude ambiguous or excessively small objects.
  • SODA-A: Targets aerial scenes and contains 2,513 high-resolution images with 872,069 instances of nine categories (airplane, helicopter, small-vehicle, large-vehicle, ship, container, storage-tank, swimming-pool, and windmill) annotated with oriented bounding boxes. Key features include large density variation, various object orientations, and diverse locations.

6. Experiments and Results:

  • Evaluation Protocol: Uses Average Precision (AP) as the evaluation metric, with a focus on the AP of small objects.
  • Implementation Details: Explains the implementation of the experiments, including the use of the mmdetection and mmrotate toolboxes, cropping and resizing images, and training details such as batch size, optimizers, and augmentation techniques.
  • Results Analysis on SODA-D:
    • Evaluates the performance of 12 representative object detection methods on the SODA-D test set.
    • Analyzes category-wise results and identifies categories with lower AP.
    • Investigates the performance of baseline detectors with different backbone networks (ResNet-50, ResNet-101, Swin-T, ConvNext-T).
    • Provides qualitative results with visualized detections.
    • Explores the impact of different label assignment strategies and loss functions.
  • Results Analysis on SODA-A:
    • Evaluates the performance of nine representative oriented object detection methods on the SODA-A test set.
    • Analyzes category-wise results.
    • Investigates the performance of baseline detectors with different backbone networks.
    • Provides qualitative results with visualized detections.
    • Explores the impact of proposal number settings for the final performance.

7. Conclusion and Future Directions:

  • The paper summarizes the key findings and contributions.
  • It highlights the need for effective feature extractors, high-quality hierarchical representations, optimized label assignment strategies, and proper evaluation metrics for SOD.

In essence, this paper provides a valuable resource for researchers working on small object detection. It offers a comprehensive overview of the field, introduces new benchmark datasets, and provides a thorough evaluation of existing methods, paving the way for further advancements in this challenging area.