Towards-Realtime-MOT
Towards-Realtime-MOT copied to clipboard
Zhongdao
mAP is always zero
Hi, when I try to train the network with the provided data, the mAP is always zero! Is it normal? How many iterations does it take to reach the higher mAP in the training phase?
Hi, @Marjanmoodi, How do you solve the problem. I try to train on different dataset but the mAP is always zero!
@Srenpan I solved it for the provided data by author, But Yes, it was always zero on custom dataset. We can Open this issue again. @Zhongdao If we want to train this network for car tracking, how we should assign IDs to different cars? Same car in different frames of one video should have same ID?
@Marjanmoodi Yes, the same car in different frames of one video should have the same ID. If the same car appears in two different videos, it should be assigned with the same ID as well.
If you train on a vehicle dataset, you had better change the anchor ratios in the config file. You may also need to tune the learning rate, batch size and other hyperparameters. A large batch size seems to be better.
Please also check #68, @Srenpan is also working on training vehicle dataset and has made some progress.
i have same problem. when i use the provided data by author, the map and p,r always 0
================================================================================ Image Total P R mAP 4 30353 0 0 0 0.259s 164 30353 0 0 0 0.149s 324 30353 0 0 0 0.15s 484 30353 0 0 0 0.149s 644 30353 0 0 0 0.153s 804 30353 0 0 0 0.15s 964 30353 0 0 0 0.15s 1124 30353 0 0 0 0.218s 1284 30353 0 0 0 0.149s 1444 30353 0 0 0 0.195s 1604 30353 0 0 0 0.161s 1764 30353 0 0 0 0.153s 1924 30353 0 0 0 0.154s 2084 30353 0 0 0 0.155s 2244 30353 0 0 0 0.152s 2404 30353 0 0 0 0.153s 2564 30353 0 0 0 0.154s 2724 30353 0 0 0 0.155s 2884 30353 0 0 0 0.155s 3044 30353 0 0 0 0.165s 3204 30353 0 0 0 0.209s 3364 30353 0 0 0 0.152s 3524 30353 0 0 0 0.16s 3684 30353 0 0 0 0.155s 3844 30353 0 0 0 0.152s 4004 30353 0 0 0 0.149s 4164 30353 0 0 0 0.155s 4324 30353 0 0 0 0.16s 4484 30353 0 0 0 0.162s 4644 30353 0 0 0 0.156s 4804 30353 0 0 0 0.158s 4964 30353 0 0 0 0.155s 5124 30353 0 0 0 0.152s 5284 30353 0 0 0 0.157s 5444 30353 0 0 0 0.164s 5604 30353 0 0 0 0.152s 5764 30353 0 0 0 0.163s
like this @Marjanmoodi How do you solve this problem in authors dataset?
alot thanks.
@Srenpan Whould you mind share your weight? or network config, I cannot to make it work to Detrac dataset even after I change the ratio of the anchors. training evaluation mAp is always zero
I solve the problem my self indeed the problem is my dataset way too small, I train to subset of DETRAC dataset. After I train to whole data its work great, even from the first epoch it is able to detect and track some vehicle... the anchor ration is 1:1
@Marjanmoodi @kikirizki I'm training the network with the provided data, (Caltech),but meet some difficulties. I Changed the code in ccmcpe.json as follows, my Caltech file in F disk has two folders, images and labels_with_ids.
{
"root":"F:/study/caltech-pedestrian-dataset-converter-master/Caltech",
"train":
{
"caltech":"./data/caltech.train"
},
"test_emb":
{
"caltech":"./data/caltech.10k.val"
},
"test":
{
"caltech":"./data/caltech.val"
}
}
But it didn't work. "Edit cfg/ccmcpe.json, config the training/validation combinations. A dataset is represented by an image list, please see data/*.train for example." I guess I do something wrong in this step.Can you share the your ccmcpe.json? By the way, can you tell me where to change the anchor ration to 1:1, because I also want to train custom dataset. Thank you!!!
To chnge anchor ratio, You can change the number right after anchors, in yolo.cfg file for example change anchors = 8, 24, .. To anchors = 24, 24 ..
@kikirizki hi，sorry to disturb you.I wonder whether you sent me an email, because I haven't received it yet.
Hi @dongziqi001 I am sorry for your waiting I have sent you the email 5 minutes ago 19:15 (gtc+7), please check your email
Hope it will help
Im sorry @dongziqi001 I sent you wrong cfg I have resent you the right cfg
hello @kikirizki ，how do you tansform the UA-DETRAC dataset to the format in JDE？I use the code in FairMOT(https://github.com/ifzhang/FairMOT/blob/master/src/gen_labels_15.py) and generate the label txts. The IDs in each txt file of each sequence like:
However, the IDs in all sequences start from 0. so in the dataloder, the total IDs=392, its Obviously wrong. can you tell me how do you process the dataset? thank you!
@kikirizki Could you share the source code for me(Or cfg and weights)? I cannot to make it work to Detrac dataset !! Please contact me Email : [email protected] Many thanks!
Hello @xdg2016 sorry for the late reply, here is my code to translate from DETRAC dataset to caltech pedestrian dataset https://github.com/kikirizki/DETRAC_dataset2Caltech_dataset , hope it will help
@Srenpan I solved it for the provided data by author, But Yes, it was always zero on custom dataset. We can Open this issue again. @Zhongdao If we want to train this network for car tracking, how we should assign IDs to different cars? Same car in different frames of one video should have same ID?
Hi, @Marjanmoodi ,
How did you solve it for the provided data?
I have also met with this on the provided data. After the first epoch, AP is always 0 during validation.
Just as https://github.com/Zhongdao/Towards-Realtime-MOT/issues/60#issuecomment-574577639
Could you please provide some suggestions?
@Srenpan I solved it for the provided data by author, But Yes, it was always zero on custom dataset. We can Open this issue again. @Zhongdao If we want to train this network for car tracking, how we should assign IDs to different cars? Same car in different frames of one video should have same ID?
Hi, @Marjanmoodi , How did you solve it for the provided data? I have also met with this on the provided data. After the first epoch, AP is always 0 during validation. Just as #60 (comment) Could you please provide some suggestions?
Sorry to bother you.
I have got the validation results which seems normal after 9th epoch.
But I still want to know why P, R, AP are all zero after 1st epoch, which seems wierd. (It could be very low, but it is supposed to have some valid values then.)
@kikirizki Sorry to bother you. i train the model using the whole dataset provided by author.But i got the low MOTA ?How to improve it? thanks
Hii @IwenLeeO Actually i did not train my model to the whole dataset provided by author, I trained my model to detract dataset instead, and I modified the anchors a little bit. But as rule of thumb make sure your configurations are the same with author has on his paper. And probably @Zhongdao can answer your question
Hi @kikirizki Thanks you for sharing.
hello,@kikirizki I am working with MOT now, Can you send me your ccmcpe.json and cfg please?Thank you very much.Please Contact me with [email protected]
Hi @kikirizki Thanks you for sharing.
hello,@kikirizki I am working with MOT now, Can you send me your ccmcpe.json and cfg please?Thank you very much.Please Contact me with [email protected]
Sure I will send you my ccmcpe.json and cfg file
Hi @kikirizki Thanks you for sharing.
hello,@kikirizki I am working with MOT now, Can you send me your ccmcpe.json and cfg please?Thank you very much.Please Contact me with [email protected]
Sure I will send you my ccmcpe.json and cfg file Hi ,@kikirizki I haven't received your email. [email protected] is mine.
Hi @kikirizki Thanks you for sharing.
hello,@kikirizki I am working with MOT now, Can you send me your ccmcpe.json and cfg please?Thank you very much.Please Contact me with [email protected]
Sure I will send you my ccmcpe.json and cfg file Hi ,@kikirizki I haven't received your email. [email protected] is mine.
@zzfpython I sent you the file you need to your email, sorry for late reply
@kikirizki Hello, sorry to bother you. I'm tracking cars in visdrone dataset, may I ask you what batch size and lr you chose? My mAP is always 0 and it does not work even if I changed the anchor ratio to 1:1...
