Avoiding Over-Detection: Towards Combined Object Detection and Counting

Abstract

Further information on publisher's website:https://doi.org/10.1007/978-3-319-59063-9 7Publisher's copyright statement:The nal publication is available at Springer via https://doi.org/10.1007/978-3-319-59063-97Additional information: Use policyThe full-text may be used and/or reproduced, and given to third parties in any format or medium, without prior permission or charge, for personal research or study, educational, or not-for-pro t purposes provided that:• a full bibliographic reference is made to the original source • a link is made to the metadata record in DRO • the full-text is not changed in any way The full-text must not be sold in any format or medium without the formal permission of the copyright holders.Please consult the full DRO policy for further details. Abstract. Existing object detection frameworks in the deep learning field generally over-detect objects, and use non-maximum suppression (NMS) to filter out excess detections, leaving one bounding box per object. This works well so long as the ground-truth bounding boxes do not overlap heavily, as would be the case with objects that partially occlude each other, or are packed densely together. In these cases it would be beneficial, and more elegant, to have a fully end-to-end system that outputs the correct number of objects without requiring a separate NMS stage. In this paper we discuss the challenges involved in solving this problem, and demonstrate preliminary results from a prototype system.

Combining Mathematical Morphology and the Hilbert Transform for Fully Automatic Nuclei Detection in Fluorescence Microscopy

“…boguslaw.obara@durham.ac.uk arXiv:1904.05217v1 [q-bio.QM] 10 Apr 2019 variations upon thresholding-based segmentation, which are limited in performance due to intensity inhomogeneity and nuclei/cell clustering [5]; -H-minima transforms [8]; voting-based techniques [18], which both show good results but are sensitive to parameters; gradient vector flow tracking and thresholding [9,10], such PDE-based methods require strong stopping and reinitialisation criteria, set in advance, to achieve smooth curves for tracking; using Laplacian of Gaussian filters [17], which is low computational complexity but struggles with variation in size, shape and rotation of objects within an image; and graph-cut optimisation approaches [12], which requires the finding of initial seed points for each nucleus. convolutional neural networks, which generally require copious amounts of manually labelled training data, and struggle to separate overlapping objects [7] In our experience, nuclei size (scale) is an important contributor to 1. whether or not an algorithm is successful 2. how robust an algorithm is to image variation and 3. to the running time of an algorithm. We have also found that many excellent algorithms exist for detecting small blobs, on the scale of a few pixels diameter, that fail or are less reliable for medium or large blobs, i.e.…”

Multi-Class Object Detection Using Adaptive Non-Maximum Suppression in Dense Images

“…Non-Maximum Suppression (NMS) is an algorithm used in the post-processing of object detection [12,14]. NMS performs well in images where objects do not overlap much, but the performance gets poor when dense object scenes or objects are overlapped and only partially visible [4,[8][9][10]18]. Therefore, to increase the detection capability of high dense overlapping objects, the optimum IoU and Confidence Score thresholds must be [7,11,19,21].…”