search

realgump / MvMHAT

MvMHAT: Self-supervised Multi-view Multi-Human Association and Tracking (ACM MM 2021, Oral Paper)
41 stars 8 forks source link

result display nan. what is the reason ? #6

Closed ssbilakeri closed 6 months ago

ssbilakeri commented 2 years ago

[2352 rows x 7 columns])]) 03:21:33 INFO - Comparing 3... 03:21:33 INFO - Comparing 2... 03:21:33 INFO - Comparing 4... 03:21:33 INFO - Comparing 1... 03:21:33 INFO - Running metrics 03:21:33 INFO - partials: 0.081 seconds. 03:21:33 INFO - mergeOverall: 0.087 seconds. IDF1 IDP IDR Rcll Prcn GT MT PT ML FP FN IDs FM MOTA MOTP IDt IDa IDm 3 0.0% 0.0% nan% nan% 0.0% 0 0 0 0 2394 0 0 0 -inf% nan 0 0 0 2 0.0% 0.0% nan% nan% 0.0% 0 0 0 0 2553 0 0 0 -inf% nan 0 0 0 4 0.0% 0.0% nan% nan% 0.0% 0 0 0 0 2847 0 0 0 -inf% nan 0 0 0 1 0.0% 0.0% nan% nan% 0.0% 0 0 0 0 2352 0 0 0 -inf% nan 0 0 0 OVERALL 0.0% 0.0% nan% nan% 0.0% 0 0 0 0 10146 0 0 0 -inf% nan 0 0 0 03:21:33 INFO - Completed

jahongir7174 commented 2 years ago

@realgump Thanks for sharing your wonderful work. I trained the model using provided dataset and tested only on the dataset "1". It seems that generated .txt files are wrong. Can you check whether it is correct or not? 1_1.txt 1_2.txt 1_3.txt 1_4.txt

realgump commented 2 years ago

Hi. The format of the .txt files seems right. You only need to check the first 6 numbers of each line is "frame_id, person_id, box_l, box_t, box_w, box_h", which is similar to the format of MOT challenge. However, your performance might be poor, since the person_id in different views seems not match. Please check if you have correctly train and inference the model.

jahongir7174 commented 2 years ago

@realgump Thanks for your quick reply. The steps that I did for training and testing are as follows:

  1. Downloaded the dataset using provided OneDrive link
  2. Changed source code for annotation loading in line number (92)[https://github.com/realgump/MvMHAT/blob/1f71d11e1cf205ef1b110150689af1055b354959/loader.py#L92] from 'gt.txt' to 'anno.txt'. There is only anno.txt in the folder, not gt.txt
  3. Removed dataset number 6 from line number (2)[https://github.com/realgump/MvMHAT/blob/1f71d11e1cf205ef1b110150689af1055b354959/config.py#L2]. There is an error while loading anno.txt for this dataset. Now, my train dataset is TRAIN_DATASET = ['1', '2', '3', '4', '5', '7', '8', '9', '10', '12', '13', '14']
  4. Changed source code for annotation loading in line number (94)[https://github.com/realgump/MvMHAT/blob/1f71d11e1cf205ef1b110150689af1055b354959/loader.py#L94] from 'det.txt' to 'anno.txt'. There is only anno.txt in the folder, not det.txt All other parts of the source code remain unchanged. Can you verify whether my steps are correct or not?
realgump commented 2 years ago

It seems correct, but I didn't met similar problem. Is your loss decrease normmaly when training？

jahongir7174 commented 2 years ago

@realgump Thanks for your reply. The loss value is decreased to 0.2890546538819386 Is it the final correct value?

jahongir7174 commented 2 years ago

@realgump Can you provide the trained weights used on the paper?