[Paper Review] Improvement of Classification in One-Stage Detector

Kehe Wu, Zuge Chen+2 more|arXiv (Cornell University)|Nov 20, 2020

Advanced Neural Network Applications26 references4 citations

TL;DR

This paper proposes an object confidence task to address misalignment between classification and localization in one-stage detectors like RetinaNet. By sharing features with classification and using IoU as targets for positive samples, the method improves classification accuracy and NMS performance, achieving +0.7% AP on COCO with ResNet50 and +1.0% with ResNet101 under identical training settings.

ABSTRACT

RetinaNet proposed Focal Loss for classification task and improved one-stage detectors greatly. However, there is still a gap between it and two-stage detectors. We analyze the prediction of RetinaNet and find that the misalignment of classification and localization is the main factor. Most of predicted boxes, whose IoU with ground-truth boxes are greater than 0.5, while their classification scores are lower than 0.5, which shows that the classification task still needs to be optimized. In this paper we proposed an object confidence task for this problem, and it shares features with classification task. This task uses IoUs between samples and ground-truth boxes as targets, and it only uses losses of positive samples in training, which can increase loss weight of positive samples in classification task training. Also the joint of classification score and object confidence will be used to guide NMS. Our method can not only improve classification task, but also ease misalignment of classification and localization. To evaluate the effectiveness of this method, we show our experiments on MS COCO 2017 dataset. Without whistles and bells, our method can improve AP by 0.7% and 1.0% on COCO validation dataset with ResNet50 and ResNet101 respectively at same training configs, and it can achieve 38.4% AP with two times training time. Code is at: http://github.com/chenzuge1/RetinaNet-Conf.git.

Motivation & Objective

• To address the persistent performance gap between one-stage and two-stage detectors despite Focal Loss in RetinaNet.
• To investigate the root cause of misalignment between high IoU localization and low classification scores in RetinaNet.
• To improve classification performance by introducing a new auxiliary task that shares features with classification.
• To reduce misalignment between localization and classification through joint use of classification scores and object confidence in NMS.
• To achieve better detection accuracy without architectural changes or additional training bells and whistles.

Proposed method

• Introduce an object confidence task that predicts IoU between anchor boxes and ground-truth boxes using shared features with the classification head.
• Use only positive samples (IoU > 0.5) to compute the object confidence loss, increasing the loss weight for high-quality positive samples.
• Train the object confidence head jointly with the classification head using a combined loss objective.
• Use the product of classification score and object confidence as the final score for Non-Maximum Suppression (NMS).
• Apply the same training configuration as RetinaNet, without data augmentation or hyperparameter tuning.
• Use the same backbone networks (ResNet50 and ResNet101) for fair comparison on MS COCO 2017.

Experimental results

Research questions

• RQ1 Why does RetinaNet still underperform compared to two-stage detectors despite using Focal Loss?
• RQ2 To what extent does misalignment between localization and classification degrade one-stage detector performance?
• RQ3 Can introducing an auxiliary object confidence task that predicts IoU improve classification accuracy and reduce misalignment?
• RQ4 Does combining classification score and object confidence improve NMS quality and final AP?
• RQ5 Can this method achieve performance gains without architectural changes or additional training tricks?

Key findings

• The method improves AP by 0.7% on MS COCO 2017 validation set when using ResNet50 under identical training configurations.
• The method achieves a 1.0% AP gain when using ResNet101 under the same training settings as baseline RetinaNet.
• The object confidence head effectively reduces misalignment between high IoU boxes and low classification scores.
• The joint use of classification score and object confidence in NMS leads to better detection box selection.
• The method achieves 38.4% AP on COCO with double the training time, demonstrating strong scalability.
• The improvements are achieved without architectural modifications or additional data augmentation.