Probabilistic Anchor Assignment with IoU Prediction for Object Detection (2007.08103v2)

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.

Overview of "Probabilistic Anchor Assignment with IoU Prediction for Object Detection"

The paper "Probabilistic Anchor Assignment with IoU Prediction for Object Detection" identifies anchor assignment as a key component in object detection systems and proposes an innovative probabilistic approach to refine this process. Traditional object detection systems often rely on the Intersection-over-Union (IoU) threshold-based heuristics for anchor assignment, which can lead to suboptimal model performance due to the simplicity of the threshold mechanism. The authors propose a probabilistic anchor assignment (PAA) strategy that uses probabilistic modeling to better distinguish positive and negative anchor samples, thus enhancing model training and subsequent detection accuracy.

Key Contributions

1. Probabilistic Model for Anchor Scores:
   • The authors introduce a novel scoring method that integrates classification and localization losses to calculate anchor scores. They fit these scores to a Gaussian Mixture Model (GMM) with two modalities (positive and negative components), enabling a probabilistic distinction between anchors.
   • Anchor probabilities derived from the GMM are used to separate positive and negative anchors, which provides more flexible and adaptive anchor assignments unlike fixed thresholds or hard IoU cutoff values commonly used.
2. IoU Prediction for Improved Scoring:
   • To address discrepancies between training and testing phase objectives in traditional object detection systems, the authors incorporate Intersection-over-Union (IoU) prediction as a measure of localization quality. IoU predictions allow the integration of localization quality into the non-maximum suppression (NMS) ranking metric, combining classification and localization for decision making during inference.
3. Score Voting for Post-processing:
   • An additional post-processing technique named score voting is proposed by the authors. It refines bounding box predictions by considering the scores of nearby boxes, further leveraging the unified classification and localization measure to improve detection performance.

Numerical Results and Experimental Validation

The experimental outcomes validate the efficacy of the proposed methods. On the MS COCO dataset, models equipped with the proposed PAA and IoU prediction mechanisms set new state-of-the-art results for single-stage detectors. The paper reports significant performance improvements over baselines and competing methods: achieving 44.8% AP on COCO test-dev with a ResNet101 backbone and 53.5% when employing a ResNeXt-32x8d-152-DCN backbone with multi-scale testing. Experiments indicate that the PAA, coupled with IoU prediction and score voting, not only enhance precision but lead to better utilization of computational resources due to the efficient assignment strategy.

Implications and Future Directions

This research has several theoretical and practical implications for the field of object detection. The probabilistic treatment of anchor scores aligns with modern machine learning principles by adopting probabilistic inference, thereby enhancing model training dynamics and performance. The integration of localization quality into object detection via IoU prediction aligns model objectives across training and testing phases, addressing one of the critical shortcomings in conventional object detection pipelines.

Looking forward, the proposed probabilistic anchor assignment framework could be further extended to more complex scenarios and tasks beyond regular object detection, such as instance segmentation and object tracking. Moreover, further exploration into the combination of probabilistic models with emerging deep learning frameworks might yield additional insights and improvements in detection systems.

In conclusion, this paper introduces a sophisticated anchor assignment strategy that integrates well within existing detection paradigms, simultaneously driving performance improvements and laying a foundation for future research endeavors.