Non-Maximum Suppression, Faster-Rcnn Network, Region Proposal Network and The Mask R-CNN (Book Report Sample)
This is a technical report where the task is to read 4 papers in the field of target detection and summarize the core of them into a report. The assignment includes Non-Maximum suppression (NMS), Faster-Rcnn Network, RPN (Region Proposal Network) network and Detailed solution of the Mask R-CNN. This assignment is very complete including source code interpretation and paper interpretation. It received a perfect score from the instructorsource..
Part I Non-Maximum suppression (NMS)
The output of the detection task is the box (including its coordinates and confidence) and its corresponding class.The NMS acts to remove the redundant box.It belongs to the post-processing link of the algorithm, generally designed manually, and is a greedy cluster based on a fixed threshold, namely, it greedily selects the highest scoring detection results and removes those neighbors beyond the threshold.
The basic algorithm process is:
For the list B of the Bounding box and its corresponding confidence S, calculate as follows:
* First, sort S in descending order, traverse all box, starting with high confidence
Ї Other box with larger box for b o U and these overlap box delete to save the current box into the result list
Ї Until all the original list are empty
i = torchvision.ops.boxes.nms(boxes, scores, iou_thres)
The Center-based detection method is a Gaussian heat map to determine the object center to do the bbox.The center point of each object represents an object, which does not need to do any NMS, and can certainly be done at the end, but with little gain.
the following is HYPERLINK "https://github.com/open-mmlab/mmdetection3d/blob/ed81c6cc326a33f927d4455c219bb7715591b08b/mmdet3d/core/post_processing/box3d_nms.py" \h mmdetection3d/box3d_nms.py at ed81c6cc326a33f927d4455c219bb7715591b08b В· open-mmlab/mmdetection3d В· GitHubThe Circular NMS code is provided, and this link also has a normal NMS code implementation.