[论文解读] Augmentation for small object detection

Máté Kisantal, Zbigniew Wojna+3 more|arXiv (Cornell University)|Feb 19, 2019

Advanced Neural Network Applications参考文献 24被引用 62

一句话总结

论文指出，在 MS COCO 上，Mask R-CNN 的小对象与锚点匹配较差，并提出对小对象进行过采样并粘贴式增强以提高小对象检测和分割，在与 Mask R-CNN 相比取得显著提升。

ABSTRACT

In recent years, object detection has experienced impressive progress. Despite these improvements, there is still a significant gap in the performance between the detection of small and large objects. We analyze the current state-of-the-art model, Mask-RCNN, on a challenging dataset, MS COCO. We show that the overlap between small ground-truth objects and the predicted anchors is much lower than the expected IoU threshold. We conjecture this is due to two factors; (1) only a few images are containing small objects, and (2) small objects do not appear enough even within each image containing them. We thus propose to oversample those images with small objects and augment each of those images by copy-pasting small objects many times. It allows us to trade off the quality of the detector on large objects with that on small objects. We evaluate different pasting augmentation strategies, and ultimately, we achieve 9.7\% relative improvement on the instance segmentation and 7.1\% on the object detection of small objects, compared to the current state of the art method on MS COCO.

研究动机与目标

• Explain why small object detection lags behind larger objects in MS COCO with Mask R-CNN.
• Propose oversampling and copy-paste augmentation to boost small-object performance.
• Evaluate augmentation strategies and quantify trade-offs between small and large object performance.

提出的方法

• Analyze anchor matching and small-object coverage in Mask R-CNN on MS COCO.
• Oversample images containing small objects during training.
• Augment images by copy-pasting small object instances with random transformations.
• Test different copy-paste strategies and pasting algorithms to assess impact on AP for small objects.
• Compare small-object performance against the baseline Mask R-CNN on MS COCO.

实验结果

研究问题

• RQ1 How does anchor-based matching affect small object detection in Mask R-CNN on MS COCO?
• RQ2 What is the impact of oversampling images with small objects on small-object AP (segmentation and detection)?
• RQ3 Can copy-paste augmentation of small objects improve small-object AP without severely harming large-object AP?
• RQ4 Which copy-paste strategies and pasting rules maximize small-object AP?

主要发现

Setting	Segmentation small AP	Segmentation medium AP	Segmentation large AP	Segmentation all AP	Detection small AP	Detection medium AP	Detection large AP	Detection all AP
Baseline	0.113	0.300	0.418	0.280	0.167	0.329	0.393	0.303
Oversampling 2x	0.120	0.299	0.409	0.279	0.173	0.328	0.387	0.304
Oversampling 3x	0.123	0.300	0.404	0.279	0.177	0.329	0.382	0.305
Oversampling 4x	0.120	0.299	0.398	0.276	0.174	0.329	0.374	0.302

• Oversampling images with small objects yields a small-object AP improvement (up to 1 percentage point in segmentation AP, ~8.85% relative gain).
• Copy-paste augmentation of small objects increases small-object AP and overall AP when combined with oversampling, with best results from a mix of original and augmented images at a 2:1 ratio.
• Copy-pasting a single small object a few times (1–3x) often yields better small-object gains than pasting many copies of all small objects.
• Avoid overlapping pasted objects and do not apply edge blur; non-overlapping pastes preserve small-object gains.
• Best overall results arise from combining oversampling with moderate augmentation (original+aug) at p=0.5 and original:augmented small-object ratio 2:1.
• The augmentation strategy yields a trade-off: improvements for small objects may come at the expense of large-object performance.