Learning Non-maximum Suppression

By Jan Hendrik Hosang, Rodrigo Benenson, Bernt Schiele. 2017

633 citations recorded by the source corpus.

Object detectors have hugely profited from moving towards an end-to-end learning paradigm: proposals, fea tures, and the classifier becoming one neural network improved results two-fold on general object detection. One indispensable component is non-maximum suppression (NMS), a post-processing algorithm responsible for merging all detections that belong to the same object. The de facto standard NMS algorithm is still fully hand-crafted, suspiciously simple, and — being based on greedy clustering with a fixed distance threshold — forces a trade-off between recall and precision. We propose a new network architecture designed to perform NMS, using only boxes and their score. We report experiments for person detection on PETS and for general object categories on the COCO dataset. Our approach shows promise providing improved localization and occlusion handling.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

End-to-End Object Detection with Transformers (2020, background): “Learnable NMS methods [16,4] and relation networks [17] explicitly model relations between different predictions with attention.”
Relation Networks for Object Detection (2017, background): “Recently, GossipNet [26] attempts to learn duplicate removal by processing a set of objects as a whole, therefore sharing the similar spirit of ours.”
Objects as Points (2019, background): “This post-processing is hard to differentiate and train [23], hence most current detectors are not end-to-end trainable.”
Panoptic Segmentation (2018, background): “To produce algorithm outputs for PS, we start from the non-overlapping instance segments from the NMS-like procedure described previously.”
DeepSetNet: Predicting Sets with Deep Neural Networks (2016, methodology): “Recently, several alternatives have been proposed to replace the greedy NMS procedure, including sequential head detection with LSTMs [18], a global optimisation approach to NMS [34, 23], or even learning NMS end-to-end using CNNs [19].”
Occluded Video Instance Segmentation: A Benchmark (2021, methodology): “In addition, some learnable NMS methods [29,53] have also been proposed, and many new methods [9,20] based on set prediction even do not need NMS post-processing.”
Acquisition of Localization Confidence for Accurate Object Detection (2018, methodology): “[11] proposes the GossipNet, a post-processing network trained for NMS based on bounding boxes and the classification confidence.”
MiniVLM: A Smaller and Faster Vision-Language Model (2020, neutral): “[56] or [12, 48, 3] to remove NMS.”
Learning to Count Objects in Natural Images for Visual Question Answering (2018, background): “Various differentiable variants such as by Azadi et al. (2017), Hosang et al. (2017), and Henderson & Ferrari (2017) exist.”
Libra R-CNN: Towards Balanced Learning for Object Detection (2019, methodology): “Other methods focus on cascade procedures [24], duplicate removal [14, 13], multi-scales [2, 1, 31, 30], adversarial learning [37] and more contextual [36].”
Bottom-Up Object Detection by Grouping Extreme and Center Points (2019, background): “A more sophisticated false-positive removal algorithm, e.g., learn NMS [18], might yield a slightly better result.”
Detection in Crowded Scenes: One Proposal, Multiple Predictions (2020, methodology): “Thus, more complex mechanisms may be introduced, for example, [18, 25] uses a neural network for more sophisticated and data-dependent duplicate removal.”
Repulsion Loss: Detecting Pedestrians in a Crowd (2017, background): “[12] works on improving the robustness of NMS, but it ends up relying on an additional network for post-processing.”
Learn to Predict Sets Using Feed-Forward Neural Networks (2020, background): “Few works [25], [26], [27] have also attempted to learn this process.”
YOLO-based Object Detection Models: A Review and its Applications (2024, positive): “NMS [20] is an efficient method used in most defect detection systems to overcome the multiple bounding box problem.”
Adaptive NMS: Refining Pedestrian Detection in a Crowd (2019, background): “Recently, soft-NMS [1] and learning NMS [14] are proposed to improve NMS results.”
Recent Advances in Deep Learning for Object Detection (2019, methodology): “Hosong et al.[166] introduced a network architecture designed to perform NMS based on conﬁdence scores and bounding boxes, which was optimized separately from detector training in a supervised way.”
YOLOv10: Real-Time End-to-End Object Detection (2024, background): “Learnable NMS [24] and relation networks [26] present another network to remove duplicated predictions for detectors.”
Object Detection in 20 Years: A Survey (2019, neutral): “Learning-Based NMS: A recent group of NMS improvements that have recently received much attention is learning-based NMS [85], [93], [109], [110], [111], [122].”