How to converting Posetrack format GT to be the one in this project?

[zhenzhiwang]

Hi brjathu,

Thanks for sharing your great work! I am wondering how you convert the original Posetrack groundtruth to be the one you used in this repo? I want to train this method but find no suitable groundtruth for the training set. I will really appreciate if you could provide the script for converting the GT format. Thank you again.

[brjathu]

Hi, Thank you, I assume you are asking about how to create posetrack_gt_data.pickle for train and validation set? Here is the code we used for creating those pickle files. Hope this helps?

def make_groundtruth_posetrack(train=False):

    ### folder structure
    # _DATA/posetrack
    # _DATA/posetrack/train
    # _DATA/posetrack/val
    # _DATA/posetrack/annotations
    # _DATA/posetrack/annotations/train
    # _DATA/posetrack/annotations/val

    if(train):
        videos_seq = "_DATA/datasets/posetrack/list_videos_train.npy"
        subfolder = "train/"
    else:
        videos_seq = "_DATA/datasets/posetrack/list_videos_val.npy"
        subfolder = "val/"

    videos_seq = np.load(videos_seq)
    video_data = {}
    posetrack_home = os.path.expanduser('_DATA/posetrack')

    for video in videos_seq:
        video_data[video] = {}
        posetrack_annotations = os.path.join(posetrack_home, 'annotations')
        coco = COCO(os.path.join(posetrack_annotations, subfolder + video + '.json'))

        img_ids = coco.getImgIds()
        imgs = coco.loadImgs(img_ids)

        posetrack_images = []
        for img in imgs:
            if not img['is_labeled']: pass
            else: posetrack_images.append(img)
        for selected_im in posetrack_images:
            ann_ids = coco.getAnnIds(imgIds=selected_im['id'])
            anns = coco.loadAnns(ann_ids)
            frame_name_x  = []; gt_idx = []; gt_bbox = []
            for ann_x in anns:
                try:
                    if(ann_x['bbox'][2]>0 and ann_x['bbox'][3]>0):
                        frame_name_x.append(selected_im['file_name'][11:])
                        gt_idx.append(ann_x['track_id'])
                        gt_bbox.append(ann_x['bbox'])
                except:
                    pass

            video_data[video][selected_im['file_name'][11:].split("/")[-1]] = [frame_name_x, gt_idx, gt_bbox]

        unique_gt_ids = []
        for frame in video_data[video]:
            bad_gt_ids = video_data[video][frame][1]
            for ix in bad_gt_ids:
                unique_gt_ids.append(ix)

        unique_gt_ids = np.unique(np.array(unique_gt_ids))

        new_ids = np.array(list(range(len(unique_gt_ids))))
        for frame in video_data[video]:
            bad_gt_ids = video_data[video][frame][1]
            new_ids = []
            for ix in bad_gt_ids:
                idx = np.where(np.array(unique_gt_ids)==ix)[0]
                new_ids.append(idx[0])

            video_data[video][frame].append(new_ids)

    if(train):
        joblib.dump(video_data, '_DATA/datasets/posetrack/gt_data_train.pkl')
    else:
        joblib.dump(video_data, '_DATA/datasets/posetrack/gt_data_val.pkl')```

[zhenzhiwang]

Yes I need the script to create posetrack_gt_data.pickle file. Thank you a lot!

By the way, I really appreciate it if you could share the training code, such as the code for loading dataset and construct losses. Thank you again!

[brjathu]

Thanks, We haven't trained any neural networks with tracking data. We just optimized 4 numbers for the distance function (https://github.com/brjathu/PHALP/blob/92206ec13310dfd6dd6a8fb01c97aac8ca82ae96/deep_sort_/nn_matching.py#L86). We will release the optimization code as soon as possible.