video_analyst
video_analyst copied to clipboard
megvii-research
result problem
I use the your training code to train. The training set only uses got10k. The test results are as follows my result:AO=0.612，SR0.5=0.735,SR0.75=0.461 your result:AO=0.595,SR0.5=0.695,SR0.75=0.479 Looks like performance is very good But when I tested vot18 with this model, the result was very poor,The test results are as follows my result:ACC=0.549,ROB=0.300,EAO=0.310 your result:ACC=0.587,ROB=0.183,EAO=0.426 Do you have this problem
It is worth noting that the code for training is "sh tools/train_test-googlenet.sh" . When training is interrupted, I run "python ./main/train.py --config experiments/siamfcpp/train/siamfcpp_googlenet-trn.yaml --resume-from-file=... " to continue training
The model with EAO=0.426 is trained with multi-datasets like coco+youtubebb+lasot+got10k. We will release the multi-datasets training code in a few weeks, please keep tuned.
Can't support multi dataset training yet. I see "names: ["GOT10kDataset",] in the configuration file, but an error will be reported when I use ["GOT10kDataset","LaSOTDataset"].
You are right. The training with multi-datasets is under testing yet, the original code is written with the internal framework. So the reproduce will take some time.
ce will take some t @lzx1413 @MARMOTatZJU Hi, can u provide the specific time of releasing the training code to reproduce the VOT18 and OTB100?
I can not guarantee one certain time but we will release it this month hopefully.
@songdejia Training code is done and it is under final test (to make sure it works well after PyTorch version change) before being merged into Master branch. If you are urgent, you can check out the "dev-fulldata_hpo" branch in my fork.
@songdejia Training code is done and it is under final test (to make sure it works well after PyTorch version change) before being merged into Master branch. If you are urgent, you can check out the "dev-fulldata_hpo" branch in my fork.
Thanks. Can u provide a readme file to describe how to use full data
Hi， I alse use your training code to train. The results of VOT2018 are rather poor. The traing full datasets are COCO2017+VID+DET+LaSOT+GOT10k+TrackingNet，the model is googlenet. my result:ACC=0.573,ROB=0.332,EAO=0.294 your result:ACC=0.587,ROB=0.183,EAO=0.426 the ROB and EAO are not good. the multi-datasets like coco+youtubebb+lasot+got10k can produce the EAO=0.426 ? the paper and the programm implementation are not consistent.
and in fact you don't use youtube-BB dataset , you use TrackingNet. there is no ytb_bb.py script in your program
@lzx1413 We will release the multi-datasets training code in a few weeks, please keep tuned? Did you have finished? Thank you !