Ucf-crime Dataset

[benedictstar]

Hi! Thanks for your wonderful work!

When I try to use your method to deal with Ucf-crime Dataset,some videos in Ucf-crime seems too long and Alphapose can't tackle them in your default setting.Regarding to this issue,I modified the parameter "q_size" in Alphapose from 1024 to 100.

Have you ever dealt with this dataset?Will this operation influences the results of the model? Or do you have better solutions?

[benedictstar]

And I have another question,which pretrained Re-ID Track model do you use? Do you use the pretrained model provided by alphapose(osnet_ain_x1_0_msmt17_256x128_amsgrad_ep50_lr0.0015_coslr_b64_fb10_softmax_labsmth_flip_jitter.pth)？

When I followed your introduction to build ShanghaiTech Dataset.,I found that the poses extracted by alphapose were close to the data you provided,however,the tracked-IDs were different from your data, which seriously affected the final result.

Looking forward to your feedback! Thanks!

[orhir]

Hi,

We haven’t tried this dataset. Our method uses a sliding window approach, thus it will only be affected by the performance of the pose estimator and not the sequence total length. I propose asking for support in the AlphaPose repo for assistance in running their code.

In addition, we didn’t use a Re-ID model, we used the Poseflow tracker, also supported by AlphaPose. Maybe the different IDs are resulted from this change.

[benedictstar]

Thanks for your reply. About my first question, I have already found that the 'q_size' will not influence the final results, thanks. But about my second question, I refer to your code 'gen_data.py'

def create_command(alphapose_dir, video_filename, out_dir, is_video=False):
    command = "python scripts/demo_inference.py"

    # loop over command line argument and add to command
    for key, val in command_args.items():
        command += ' --' + key + ' ' + val
    if is_video:
        command += ' --video ' + video_filename
    else:
        command += ' --indir ' + video_filename
    command += ' --sp'  # Torch Re-ID Track
    command += ' --pose_track'  # Torch Re-ID Track
    # command += ' --detector yolox-x'  # Simple MOT Track

    return command

I find that the command you used actually calls the Re-ID Track model from Alphapose by

command += ' --sp'  # Torch Re-ID Track
command += ' --pose_track'  # Torch Re-ID Track

I also tried to call the Poseflow tracker from Alphapose to track the person, however, the file I got was not a JSON file in the format like yours. Would you please update the file 'gen_data.py' if you upload the incorrect version? Thank you!

[orhir]

Hi, Sorry for the late response. We used PoseFlow as a tracker for ShanghaiTech and UBnormal datasets. As we provide the extracted data for these datasets, the gen_data.py code for custom data/other datasets. The default here is indeed the Re-ID option, as the PoseFlow implementation is quite slow.

Can you please provide the output data you receive after using PoseFlow tracker, so we can debug this code?

[wrtnew]

Thanks for your reply. About my first question, I have already found that the 'q_size' will not influence the final results, thanks. But about my second question, I refer to your code 'gen_data.py'

def create_command(alphapose_dir, video_filename, out_dir, is_video=False):
    command = "python scripts/demo_inference.py"

    # loop over command line argument and add to command
    for key, val in command_args.items():
        command += ' --' + key + ' ' + val
    if is_video:
        command += ' --video ' + video_filename
    else:
        command += ' --indir ' + video_filename
    command += ' --sp'  # Torch Re-ID Track
    command += ' --pose_track'  # Torch Re-ID Track
    # command += ' --detector yolox-x'  # Simple MOT Track

    return command

I find that the command you used actually calls the Re-ID Track model from Alphapose by

command += ' --sp'  # Torch Re-ID Track
command += ' --pose_track'  # Torch Re-ID Track

I also tried to call the Poseflow tracker from Alphapose to track the person, however, the file I got was not a JSON file in the format like yours. Would you please update the file 'gen_data.py' if you upload the incorrect version? Thank you!

Hi, have you already solve it?

[orhir]

Hi, Can you please provide the JSON file which is in the wrong format?

[benedictstar]

Hi, Sorry for my late reply. Actually, when I directly use Poseflow in AlphaPose to extract pose sequences from the videos, I don't get any JSON files but some other kinds of files. Then I deleted the wrong files. So I am sorry that I can't provide the wrong files right now. At that time, I just changed the command command += ' --pose_track' to command += ' --pose_flow' as AlphaPose required. Hope this can help you to solve the problems. I notice that you provided a new code for tackling the problem of extracting pose sequences using Poseflow in another issue, maybe that can help me, Thanks a lot.

[benedictstar]

I notice that you provided a new code for tackling the problem of extracting pose sequences using Poseflow in another issue, maybe that can help me, Thanks a lot.

Hi, I haven't solved the problems yet, but I notice that the author provided a new code for tackling the problem of extracting pose sequences using Poseflow in another issue #15, maybe that can help you.

[1190200311]

only PoseFlow can provide relatively accurate ids. For instance, their reID in a video with 5 people may result in 9 to 10 ids, making it difficult to match ids with scores