liuqk3 / otcd Goto Github PK

Hi Team,

Where I can find the code to train the network responsible to generate appearance descriptors for a track.

Thanks

Hi
When I try to run, I get the error No module named 'lib.model.roi_align.roi_align._ext.crop_and_resize._crop_and_resize'
Can you help?
Thank you.

Thanks for your work firstly.
I have tried to reproduce your work recently, with your training code. I forcused on the tracking model, the single head one.
But I can't get your results in the paper. My trained model only can give me the bboxes, whose left-bottom point fixed, the same as the input bbox, and right-down point moving to right and down, every time inferencing with the model.
So I want to ask for your help, that do you have any tricks in your training phase, and if training with the code without any change can get the results.

Hi Team,

I can't download the pre-trained models from the link you provided in the readme file. It looks like the URL no longer exists. Can you please provide me an updated URL to download the pre-trained models.

Thanks

Hi,
Thanks for the nice work, I have few questions about the training of the detection network. I try to run your training script of the detection network, which is train_detection_on_mot.py in useful_scripts. However, the loss printed on the console seems abnormal, in the first epoch, the rcnn's regression loss is 100x smaller than three other losses which are rpn_cls, rpn_box and rcnn_cls. How do you train the RFCN network? Does the training process of RFCN similar to Faster-RCNN with multiple steps?

Hi,
I'm getting the following error while running the tracking_on_mot.py

Called with args:
Namespace(cfg_file='./cfgs/resnet101.yml', class_agnostic=False, cuda=True, dataset_year=['MOT16'], detection_interval=1, detection_sbc_model='./save/detection_sbc_101_4_1_9417.pth', detectors=['PRIVATE'], feature_crop_size=(1024, 7, 7), im_crop_size=(3, 120, 40), im_for_box=False, iou_or_appearance='both', large_scale=False, mGPUs=False, mot_dir='MOT', mv_crop_size=(2, 120, 40), mv_for_box=False, res_for_box=False, residual_crop_size=(3, 120, 40), save_detections_with_feature=False, set_cfgs=['ANCHOR_SCALES', '[1, 4, 8, 16, 32]', 'ANCHOR_RATIOS', '[1, 3]', 'MAX_NUM_GT_BOXES', '55'], stage=['val'], tracking_box_transform_sigma=1.5, tracking_model='./save/tracking_net_single_resnet18_mv_residual_2_10_6034.pth', vis=False)
tracking on MOT16-09 using PRIVATE detector ...
 Warming up the tracker...
warming up, frame 1
Decode Error.
Decoding video failed

I have downloaded MOT16 images from the official website and converted the images into a video with 20fps and encoded that video using ffmpeg as described in pytorch-coviar repository. I believe that this could be the problem of encoder. Could you share the configuration and commands of ffmpeg encoding that you have used? And if possible, can you also share the link to the encoded video that can be used directly?