What Makes for Effective Detection Proposals?

By Jan Hendrik Hosang, Rodrigo Benenson, Piotr Doll'ar, Bernt Schiele. 2015

769 citations recorded by the source corpus.

Current top performing object detectors employ detection proposals to guide the search for objects, thereby avoiding exhaustive sliding window search across images. Despite the popularity and widespread use of detection proposals, it is unclear which trade-offs are made when using them during object detection. We provide an in-depth analysis of twelve proposal methods along with four baselines regarding proposal repeatability, ground truth annotation recall on PASCAL, ImageNet, and MS COCO, and their impact on DPM, R-CNN, and Fast R-CNN detection performance. Our analysis shows that for object detection improving proposal localisation accuracy is as important as improving recall. We introduce a novel metric, the average recall (AR), which rewards both high recall and good localisation and correlates surprisingly well with detection performance. Our findings show common strengths and weaknesses of existing methods, and provide insights and metrics for selecting and tuning proposal methods.

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

Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (2015, methodology): “In the following, we train Fast R-CNN using 2k RPN proposals, but evaluate different numbers of proposals at test-time.”
Mask R-CNN (2017, methodology): “R-CNN: The Region-based CNN (R-CNN) approach [13] to bounding-box object detection is to attend to a manageable number of candidate object regions [38, 20] and evaluate convolutional networks [25, 24] independently on each RoI.”
The Cityscapes Dataset for Semantic Urban Scene Understanding (2016, methodology): “Many of the commonly used bounding box proposal methods [28, 54] first generate a set of overlapping segments, e.”
Inside-Outside Net: Detecting Objects in Context with Skip Pooling and Recurrent Neural Networks (2015, methodology): “Using dynamic pooling [13] we can efﬁciently evaluate thousands of different candidate ROIs with a single forwards pass of the network.”
Learning to Segment Object Candidates (2015, methodology): “These methods achieve top results on object detection (when coupled with R-CNNs [10]) and also obtain the best AR [13].”
Focal Loss for Dense Object Detection (2017, background): “contain all objects while filtering out the majority of negative locations [28], and the second stage classifies the proposals into foreground classes/background.”
Object-Proposal Evaluation Protocol is ‘Gameable’ (2015, background): “In [59], authors present an analysis of various proposal methods regarding proposal repeatability, ground truth annotation recall, and their impact on detection performance.”
HyperNet: Towards Accurate Region Proposal Generation and Joint Object Detection (2016, methodology): “The detection accuracy with respect to mAP is shown in Table 5.”
R-CNN minus R (2015, methodology): “These methods, which are extensively reviewed in [13], originate in the idea of “objectness” proposed by [1].”
Attend Refine Repeat: Active Box Proposal Generation via In-Out Localization (2016, background): “Given how similar are the tasks of category-speciﬁc object detection and category-agnostic proposal generation, it is safe to assume that a similar behaviour will probably hold for the latter one as well.”