Which Faster RCNN repo do you use during testing and validation

[microsoft]

Which Faster RCNN repo do you use during testing and validation

[bearpaw]

Hi @HuAndrew , same question here.

I am playing with this code recently and was also wondering about how did you generate the detection part.

To be more specific, I am detecting humans from the COCO val 2017 keypoints images (5000 images) from the person_keypoints_val2017.json. I try to use Yolo v3 detector and keep only the bounding boxes regarding humans. Then I dump the JSON file which is similar to this repo's.

However, the size of the generated JSON is quite small compared with theirs (~1.3MB vs 16.4MB). Also, when I run cocoEval and use person_keypoints_val2017.json as groundtruth, I can only get about 40 AP.

Any suggestions? Thank you in advance :)

[yurymalkov]

I have the same question. Can you please share your detector or give a link to a similar one?

[namheegordonkim]

👍 . Related papers keep mentioning of the "person detector used in Simple Baseline..." but it's nowhere to be found

[Odaimoko]

Hi @HuAndrew , same question here.

I am playing with this code recently and was also wondering about how did you generate the detection part.

To be more specific, I am detecting humans from the COCO val 2017 keypoints images (5000 images) from the person_keypoints_val2017.json. I try to use Yolo v3 detector and keep only the bounding boxes regarding humans. Then I dump the JSON file which is similar to this repo's.

However, the size of the generated JSON is quite small compared with theirs (~1.3MB vs 16.4MB). Also, when I run cocoEval and use person_keypoints_val2017.json as groundtruth, I can only get about 40 AP.

Any suggestions? Thank you in advance :)

Well the author said 56.4 AP on person category. I have used Detectron's model . In End-to-End Faster & Mask R-CNN Baselines, the entry X-101-64x4d-FPN with 42.4 box AP can get 55.7 AP on person cat. I think this is competitive.

[HuAndrew]

@bearpaw @Odaimoko Hello, I test multi detector, like mask, cascade_RCNN , and the detector vis and other preds' results are as follows:

vis samples

preds samples

256x192_pose_resnet_50_d256d256d256	total person	detect AP	keypoint
ground truth	11004	XXXXX	72.4
faster author	104125	56.4	70.5
mask rcnn_0.7	13167	48.6	68.1
mask rcnn_0.5	15530	49.5	68.6
mask rcnn_0.3	15796	49.6	68.7
Cascade_RCNN	73597	53.0	70.0

Then
From the test results, something can be found:

• In order to achieve the purpose of rescore tricks, the author let detector gives multiple detection boxes for every person instance(rescore operation refer to COCO17-Keypoints-TeamOKS). And rescore could amend pred results.
• Then if we want to get multi bboxs, we can adjust NMS postprocess.
• As long as the detector position is very correct like gt bbox, the prediction results are also very good.
• Top-down methods, the detector is very import to improve preds results.
• But I use multi bbox, preds results are worse. So I guess the author used the byte bboxs and the NMS operation together amend the detectors performance.
• Other detectors: maskrcnn-benchmark, yolov3.

Welcome to Join pose forum www.ilovepose.com

[wmcnally]

Evaluated using the Detectron2 repo:

• Faster R-CNN with ResNeXt-101 FPN backbone gets 56.6 AP for the person category on COCO val2017.
• Faster R-CNN with ResNet-101 FPN backbone gets 55.7 AP for the person category on COCO val2017.

https://github.com/facebookresearch/detectron2/blob/main/MODEL_ZOO.md