search

tangyuhao / GOTURN-Tensorflow

A tensorflow implementation for GOTURN tracker
MIT License
130 stars 61 forks source link

Urgent:how to make train_file.txt #12

Open anjali1601 opened 6 years ago

anjali1601 commented 6 years ago

Please tell how to make this train_file.txt for ALOV and imagenet?? or it is already available?

subhadip3 commented 6 years ago

@anjali1601 did u figure it out? @tangyuhao please reply

anjali1601 commented 6 years ago

@subhadip3 No i have not yet find out how to make train_set.txt for ALOV dataset? one file(train_set.txt) is already in ZIP that is for imagenet and ALOV for both?

subhadip3 commented 6 years ago

Ok, I went along with cropping ALOV images and put them in target and searching folders for training.

anjali1601 commented 6 years ago

i didnt get it.. please tell me what is train_set.txt file? I have ALOV and imagenet dataset ..now what will be the next step to run this program?

subhadip3 commented 6 years ago

Along with ALOV images you will also find bounding box data there. Just crop the images to 2X the size of bounding boxes to obtain searching and target images(do it as described in paper). Then create your own train_set.txt file with image names(target and searching) and bounding box positions in each line and run the program

anjali1601 commented 6 years ago

thank you so much..I will look into this..Also what about imagenet (47GB) ?

subhadip3 commented 6 years ago

I am only using ALOV right now :)

anjali1601 commented 6 years ago

Thank you so much :)

anjali1601 commented 6 years ago

@subhadip3 have u successfully implemented this? I have cropped the ALOV images using linear interpolation. I am confused how to make train file: Example of one line: train/target/000024.jpg, train/searching/000024.jpg, 0.29269230769230764, 0.22233115468409587, 0.7991794871794871, 0.7608061002178649

are these target and search folders have same images? and these bounding boxes is of search image or target image? and how these bounding box will be calculated?

anjali1601 commented 6 years ago

@subhadip3 please reply

subhadip3 commented 6 years ago

@anjali1601 Sorry for the late reply. The target folder should contain crops where the object is centred and the searching folder should contain random crops such that the corresponding object is not always centred. While making these random crops note the position of the bounding boxes wrt these new crops and scale them so that they lie between 0-1. As mentioned in the paper I have made sure to only include crops that include atleast 50% of the object.

anjali1601 commented 6 years ago

@subhadip3 thank you so much for your help :)