Augmentation for small object detection (1902.07296v1)

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.

• The paper demonstrates that augmentation and oversampling strategies boost small object detection, with a 9.7% improvement in instance segmentation.
• It identifies dataset imbalance and suboptimal anchor matching as critical challenges in training Mask R-CNN for small object detection.
• Empirical results show that a balanced mix of original and augmented images enhances detection precision without degrading large object performance.

Augmentation for Small Object Detection: A Detailed Analysis

This paper presents an in-depth exploration into the challenge of detecting small objects within the field of computer vision, using the state-of-the-art framework Mask R-CNN evaluated on the MS COCO dataset. The authors address a significant issue in modern object detection systems: the disparity in performance between detecting small and large objects. The paper identifies key factors contributing to this performance gap and proposes innovative solutions, focusing primarily on data augmentation and oversampling strategies.

Core Contributions

The researchers identify two main challenges in detecting small objects:

1. Imbalance in Dataset Representation: A smaller proportion of images contain small objects, which biases models towards learning features of medium and large objects.
2. Suboptimal Anchor Matching: The intersection-over-union (IoU) between small objects and proposed anchors is frequently below the desired threshold, impeding reliable training on small objects.

To address these issues, the authors propose two primary strategies: oversampling and augmentation.

Proposed Methodologies

• Oversampling: The authors demonstrate that by increasing the frequency of images containing small objects during training, models achieve better precision on these smaller instances. They empirically evaluate different oversampling ratios (2×, 3×, 4×) and discover an optimal balance that enhances small object detection without significantly harming larger object detection performance.
• Augmentation via Copy-Pasting: Instead of traditional augmentation methods, the authors employ a copy-pasting strategy, where small objects are duplicated and inserted at varied locations within the same image. This technique enriches the dataset with greater variability of small object positioning, thus improving the model's aptitude to detect small objects.

The paper reports a relative improvement of 9.7% in instance segmentation and 7.1% in object detection for small objects by combining these strategies with the Mask R-CNN.

Experimental Insights

Several experimental configurations were tested to refine the augmentation process:

• Varying the number of objects pasted and evaluating the interaction between oversampling and augmentation revealed that an optimal strategy involved balancing original and augmented image instances.
• The paper confirmed the importance of non-overlapping pasting and noted that edge blurring techniques did not yield significant improvements, suggesting the copy-paste process should maintain original object characteristics to be effective.

Implications and Future Directions

The implications of these findings suggest that simple yet effective augmentation techniques can significantly bridge the performance gap in small object detection. Practically, methods described in this paper can enhance critical applications like autonomous vehicles and satellite imaging, where small object detection is pivotal.

Future research may extend these augmentation techniques to dynamic datasets or explore the integration with other network architectures. Additionally, further refinement in augmentation processes, perhaps with contextual understanding, could provide new directions for improving performance even further.

Overall, this paper provides a robust framework for improving small object detection in computer vision, setting a precedent for future explorations in specialized dataset augmentations.