Probabilistic Anchor Assignment with IoU Prediction for Object Detection

Abstract

In object detection, determining which anchors to assign as positive or negative samples, known as anchor assignment, has been revealed as a core procedure that can significantly affect a model's performance. In this paper we propose a novel anchor assignment strategy that adaptively separates anchors into positive and negative samples for a ground truth bounding box according to the model's learning status such that it is able to reason about the separation in a probabilistic manner. To do so we first calculate the scores of anchors conditioned on the model and fit a probability distribution to these scores. The model is then trained with anchors separated into positive and negative samples according to their probabilities. Moreover, we investigate the gap between the training and testing objectives and propose to predict the Intersection-over-Unions of detected boxes as a measure of localization quality to reduce the discrepancy. The combined score of classification and localization qualities serving as a box selection metric in non-maximum suppression well aligns with the proposed anchor assignment strategy and leads significant performance improvements. The proposed methods only add a single convolutional layer to RetinaNet baseline and does not require multiple anchors per location, so are efficient. Experimental results verify the effectiveness of the proposed methods. Especially, our models set new records for single-stage detectors on MS COCO test-dev dataset with various backbones. Code is available at https://github.com/kkhoot/PAA.

Disentangle Your Dense Object Detector

“…For fair comparison, results with single-model and single-scale testing at a short side of 800 are reported. When using the same backbone network, our DDOD wins all other methods including the recently developed GFL [19] and PAA [16]. Additionally, empowered by the advanced Res2Net-101-DCN backbone, DDOD achieves a 52.5 mAP in the single-scale testing protocol, suggesting its great prospect in various applications.…”

“…ATSS [52] suggests an adaptive anchor assignment which is calculated by the statistics from the mean and standard deviation of IoU values on a set of anchors for each GT. In order to adaptively separate anchors according to the models's learning status, PAA [16] investigates a probabilistic manner to assign labels. However, these methods do not discuss the label conjunction between the classification and regression branches.…”

H2Det: A High-Speed and High-Accurate Ship Detector in SAR Images

“…To demonstrate the superiority of the proposed method, we compare it with other state-of-the-art methods such as RepPoints [14], SSD [8], RetinaNet [9], Foveabox [13], NAS-FPN [11], GHM [39], Dynamic RCNN [40], YOLOv3 [29], CARAFE [41], Weight Standardization [42], Generalized Attentionn [43], Guided Anchoring [44], Free Anchor [45], Faster RCNN [6], Libra RCNN [12], ATSS [46], FSAF [47], PAA [48], and VFNet [15] which are implemented by object detection toolbox MMDetection. According to the descending order of AP, Table Ⅻ shows the evaluation metrics of different methods on SSDD dataset.…”

Modulating Localization and Classification for Harmonized Object Detection

“…MAL [20] selects positive anchors by combining their classification and localization scores. PAA [21] uses Gaussian Mixture Model to separate positive and negative anchors based on a carefully designed anchor score. AutoAssign [22] introduces a confidence weighting module for adaptive assignment.…”