ROLO
ROLO copied to clipboard
Guanghan
how to get the yolo output data?
@Guanghan Hello Guanghan, I want to apply your code to test in my video data. But I confused how to get the yolo output in my data from your method. Also, I found that your offered DATA /yolo_output/ *.npy file have array with 4102 dims. In this array, the lastest 6 data are the coordinates of object obtained from yolo as you said in paper. so I have question about how did you choose the object coordinate information from all yolo output objects' coordinate information.
best wishes!
He separates out the target of interest from the rest by applying IoU of all detection rectangles with ground truth values and selecting the rectangle with max IoU score. This means that even for evaluation of your test set, (if it contains multiple detections) you'd need ground-truth rectangle values. And yes, that's a disadvantage of the current approach.
@suryaprakaz Oh,It's not a good idea to use ground-truth when testing. I think @Guanghan may use other ways to locate the tracking object from all yolo output. The result may exist in 3rd party/sort_yolo.py. I will check over the python file later and give an answer.
@suryaprakaz and @dongqiaqia Could you change the YOLO_network.py in a way that could give you the results without using GT?
@suryaprakaz @dongqiaqia @Guanghan I m very confused when i read the codes. Yes, I think u r right. The approach uses ground truth of test images. For each test image if it gives a bounding box, and the bounding box is not generated by network. How can it call tracking ? I think LSTM is more like another bounding regression.... It can not be used in the video has no ground truth...
@dongqiaqia @suryaprakaz Did u read molo codes? I thought it will help some but now i have no idea about it... How can i find the "seq_dets"? I read the old issues just now. Do there have new solution about the ROLO's problem now?
@Anida-qin @dongqiaqia @masoumeh1 @suryaprakaz If you are find the solution of using the molo code, Please tell me!!!
