pva-faster-rcnn
pva-faster-rcnn copied to clipboard
sanghoon
the Small Object detection issue
Dear All,
When i training my own dataset(229 category logo, totally 3w+ images) using PVANet, i got mAP = 0.9672, indeed it is very high. PVANet is indeed so good, the original R-FCN is 0.953(i do a modify it get to 0.9705). by the way, they use different backbone. so it is just a reference. but i find that the small object detection result is not as good as faster-rcnn and R-FCN.
For the Small object, the faster-rcnn get the best result, but its false alarm is also the highest. R-FCN get middle result and the lowest false alarm. PVANet get a worse small object detection result and middle false alarm result.
So, I feel confuse about this result. because of using the idea of HyperNet, the PVANet supposed to be have a good performance in small object detection, but it is not in my dataset...
Anyone please give me some tips about this issue, How can i fix it and get a better result? @sanghoon @kyehyeon
I think small objects maybe filtered by nms during generating proposals, you may check the nms parameter
@argman Dear argman, i check the cfg file lib/fast_rcnn/config.py:
__C.TRAIN.RPN_MIN_SIZE = 16 __C.TEST.RPN_MIN_SIZE = 16
it seems the parameter is OK, So What's the problem about this issue?
dear xiaoxiongli,I also want to train my own dataset with py-R-FCN,however i cannot modify my model correctly.Maybe, there are something else i didn't know ,for example i didn't get the right pre-trained model ,could you please show me your models and give me some advice.My Email address is [email protected] 。thank you so much
@Simonhong111 i am not use py-R-FCN, i implement the 4-stage training in MATLAB according the shaoqingren's faster-rcnn MATLAB code, I think you can ask the py-R-FCN code's author for help. i think after i done the PVANet, i will release the MATLAB version R-FCN 4-stage training code.
Hi @xiaoxiongli, I guess maybe it's due to limited computational costs. Could you share some quantitative comparison results?
@sanghoon Dear sanghoon, sure, i already send you a e-mail with the test result and also some comment, thank you very much^_^...
@Caojunxu NO!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
@xiaoxiongli How small of objects do you need to detect? (16x16 is the smallest object?) I also want to resolve this problem, SSD (single shot multibox detection) also has this problem.
@ck196 above 24x24, or maybe above 32x32.................................... i send you a email and hope discuss with you when you are free
@xiaoxiongli hi , I wonder how you finetune your own data? Use full/test.model or example_finetune384/? when I use test.model,I can't get a good result.When I use exampe_finetune384/,the speed is slow(0.18s/img,gtx1070).Can u give me some advice? Thank you in advance!
@AITTSMD example_finetune384, i use K40. in the aspect of speed, make sure that you do NOT compile caffe with cudnn, because cudnn's 1x1 convolution is slow.
hello,why the caffemodel th author provided i cant download it, do you success? if you hava download them, could you share with me?thanks a lot~
@xiaoxiongli
I appreciate if you have a look at my question: #63 Also I appreciate if you inform me if you have published your MATLAB version R-FCN 4-stage training codes.
Facing the same small object not detected problem when using SSD. Was anyone able to find the cause? Did tweaking the nms parameter help? Would greatly appreciate it if someone could share their experience /solution attempts
