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

A Normalized Gaussian Wasserstein Distance for Tiny Object Detection (2110.13389v2)

Published 26 Oct 2021 in cs.CV

Abstract: Detecting tiny objects is a very challenging problem since a tiny object only contains a few pixels in size. We demonstrate that state-of-the-art detectors do not produce satisfactory results on tiny objects due to the lack of appearance information. Our key observation is that Intersection over Union (IoU) based metrics such as IoU itself and its extensions are very sensitive to the location deviation of the tiny objects, and drastically deteriorate the detection performance when used in anchor-based detectors. To alleviate this, we propose a new evaluation metric using Wasserstein distance for tiny object detection. Specifically, we first model the bounding boxes as 2D Gaussian distributions and then propose a new metric dubbed Normalized Wasserstein Distance (NWD) to compute the similarity between them by their corresponding Gaussian distributions. The proposed NWD metric can be easily embedded into the assignment, non-maximum suppression, and loss function of any anchor-based detector to replace the commonly used IoU metric. We evaluate our metric on a new dataset for tiny object detection (AI-TOD) in which the average object size is much smaller than existing object detection datasets. Extensive experiments show that, when equipped with NWD metric, our approach yields performance that is 6.7 AP points higher than a standard fine-tuning baseline, and 6.0 AP points higher than state-of-the-art competitors. Codes are available at: https://github.com/jwwangchn/NWD.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (4)
  1. Jinwang Wang (6 papers)
  2. Chang Xu (323 papers)
  3. Wen Yang (185 papers)
  4. Lei Yu (234 papers)
Citations (212)

Summary

An Analysis of "A Normalized Gaussian Wasserstein Distance for Tiny Object Detection"

The paper "A Normalized Gaussian Wasserstein Distance for Tiny Object Detection" by Jinwang Wang, Chang Xu, Wen Yang, and Lei Yu proposes a novel approach to address the challenges in detecting tiny objects—a task complicated by the limited pixel size of the objects often leading to inadequate appearance information. The authors critique the reliance on Intersection over Union (IoU) based metrics, highlighting their sensitivity to minor location deviations, particularly in the context of anchor-based detectors.

Core Contribution

The central contribution of the paper is the introduction of the Normalized Gaussian Wasserstein Distance (NWD) as an alternative metric for evaluating tiny object detection. By modeling bounding boxes as 2D Gaussian distributions and measuring similarity using the Wasserstein distance, NWD offers a scale-invariant and robust metric less sensitive to discrete location deviations than IoU. The paper thoroughly integrates NWD into the anchor-based object detection pipeline, encompassing the assignment of positive/negative labels, non-maximum suppression (NMS), and the loss functions traditionally reliant on IoU metrics.

Methodology and Results

  1. Gaussian Modeling and NWD: The paper models bounding boxes as 2D Gaussian distributions, leveraging the properties of Wasserstein distance to compute a similarity measure regardless of overlap presence. The NWD is proposed to attenuate the drawbacks associated with IoU, particularly in scenarios with minimal overlap and embedded in both single-stage and multi-stage detectors.
  2. Performance Evaluation: The research is evaluated on the AI-TOD dataset, specifically designed for tiny object detection with an average object size significantly smaller than other standard datasets. The proposed NWD-integrated detectors demonstrate a performance improvement of 6.7 AP points over baseline IoU-based methods, with a notable 6.0 AP point gain over the state-of-the-art alternatives.
  3. Extensive Ablation Studies: To differentiate the impact of NWD from traditional IoU-based metrics, various configurations were experimented with, revealing the substantial improvements NWD brings in label assignment and bounding box regression tasks. Further insights are provided through sensitivity analysis, demonstrating NWD's consistent performance across varying object scales.

Implications and Future Directions

This research offers practical enhancements for tiny object detection scenarios, crucial for applications like surveillance, autonomous driving, and maritime rescue missions. The deployment of NWD can potentially catalyze advancements in small object detection across different scales and complexities, paving the way for enhanced robust object detection frameworks. Future directions might explore extending this approach to tackle challenges in three-dimensional object detection or integrating NWD with emerging neural architectures for improved scalability and adaptability in dynamic environments. Furthermore, cross-evaluation with additional datasets would strengthen the generalizability of NWD in diverse application settings.

The normalization of Wasserstein distance not only addresses the pitfalls inherent in traditional metrics but also establishes a foundational shift towards more nuanced methods of evaluating object detection efficacy, particularly in the field of subtle and complex detection challenges posed by tiny objects.