In "DepthTrack: Unveiling the Power of RGBD Tracking" it said

RGB images were stored as 24-bit JPEG with low compression rate and the depth frames as 16-bit PNG.

I use following code to read depth png:

from PIL import Image
import numpy as np
a = np.array(Image.open('....../adapter02_indoor/depth/00000001.png'))

And I find it's dtype is int32, how to understand?

Hello,

I want to know how to generate the per attribute F-scores (like the Figure 7 in your paper) when analyse the results of trackers? Is this figure generated by using VOT tookit, and if so, how does it work? At your convenience, would you please help me with this problem?

Thanks for your assistance.

Hi! I have already trained a detector on my dataset. What I would like to do now is to run a tracker of yours on my dataset, providing in input the results from detections. For example I have succesfully run some tracking algorithms as SORT, providing in input a format file as the MOT standard input. can you guide me in how to provide my dataset in input to your method?

The following is broken lasot depth zip list

'kangaroo-10',
 'kangaroo-11',
 'kangaroo-13',
 'kangaroo-17',
 'kangaroo-20',
 'kangaroo-3',
 'kangaroo-4',
 'kangaroo-6',
 'kangaroo-8',
 'kangaroo-9',
 'lion-10',
 'lion-11',
 'lion-12',
 'lion-15',
 'lion-18',
 'lion-2',
 'lion-20',
 'lion-4',
 'lion-7',
 'lion-8',
 'lizard-10',
 'lizard-11',
 'lizard-12',
 'lizard-13',
 'lizard-14',
 'lizard-15',
 'lizard-17',
 'lizard-19',
 'lizard-3',
 'lizard-5',
 'lizard-6',
 'lizard-8',
 'lizard-9',
 'microphone-16',
 'microphone-17',
 'microphone-18',
 'microphone-19',
 'microphone-3',
 'microphone-6',
 'monkey-1',
 'monkey-10',
 'monkey-11',
 'monkey-13',
 'monkey-14',
 'monkey-15',
 'monkey-16',
 'monkey-17',
 'monkey-2',
 'monkey-20',
 'monkey-3',
 'monkey-4',
 'monkey-5',
 'monkey-6',
 'monkey-8',
 'monkey-9',
 'motorcycle-12',
 'motorcycle-14',
 'motorcycle-17',
 'motorcycle-2',
 'motorcycle-20',
 'motorcycle-3',
 'motorcycle-6',
 'motorcycle-7',
 'motorcycle-9',
 'person-1',
 'person-11',
 'person-14',
 'person-15',
 'person-17',
 'person-18',
 'person-5',
 'person-6',
 'person-9',
 'pig-1',
 'pig-10',
 'pig-11',
 'pig-12',
 'pig-13',
 'pig-14',
 'pig-15',
 'pig-16',
 'pig-17',
 'pig-18',
 'pig-19',
 'pig-5',
 'pig-6',
 'pig-7',
 'rabbit-1',
 'rabbit-10',
 'rabbit-11',
 'rabbit-13',
 'rabbit-15',
 'rabbit-16',
 'rabbit-17',
 'rabbit-2',
 'rabbit-20',
 'rabbit-4',
 'rabbit-7',
 'rabbit-8',
 'rabbit-9',
 'robot-1',
 'robot-10',
 'robot-11',
 'robot-13',
 'robot-16',
 'robot-17',
 'robot-18',
 'robot-19',
 'robot-2',
 'robot-4',
 'robot-5',
 'robot-6',
 'robot-7',
 'robot-8',
 'rubicCube-10',
 'rubicCube-11',
 'rubicCube-12',
 'rubicCube-13',
 'rubicCube-14',
 'rubicCube-15',
 'rubicCube-16',
 'rubicCube-17',
 'rubicCube-18',
 'rubicCube-19',
 'rubicCube-20',
 'rubicCube-4',
 'rubicCube-5',
 'rubicCube-8',
 'rubicCube-9'

Appreciate that your hard fix work. BR.

Hello,

It's really great that you guys have disclosed such an excellent work.

I would like to know how to obtain the F-score, precision and recall values when I get the raw results with '.txt' suffix?

Thanks a lot!

Hi~ Thanks for the great work and code!

I run the code with Kinect and found it tracks objects quite well with a user-defined bounding box. I'm wondering if I can use the DeT tracker with some detectors e.g., YOLO/MaskRCNN. For example, the detectors predict the instance segmentation with bboxes and the bboxes are tracked by the DeT. Finally, the DeT will assign these bboxes with new IDs or associate them with previous IDs. However, the new bbox tracking is defined by users in the current setting, how to filter the detections automatically?

Is there any script or a hint for doing this? Thank you very much!

depthtrack.py gt = pandas.read_csv(bb_anno_file, delimiter=',', header=None, dtype=np.float32, na_filter=False, low_memory=False).values maybe na_filter should be true?

Following sequence_list tell us there is a seq called lock_wild in depthtrack test set, but actully it's in training set link you provide and called lock01_wild. At the same time, there is a seq called lock02_indoor in test set link you provide. https://github.com/xiaozai/DeT/blob/8aa5e2eae91e0f3f88290b4133aadf09aebb34b6/pytracking/evaluation/depthtrackdataset.py#L114

Please note this.

File "../ltr/data/transforms.py", line 238, in transform_image rgb = tvisf.normalize(rgb, self.mean, self.std, self.inplace) TypeError: normalize() takes 3 positional arguments but 4 were given

Hi,

Thanks for the brilliant work on RGB-D tracking.

I am new to the tracking domain. I have several questions regarding the tracking datasets. Can you please help me to clarify several points?

Firstly, in the paper it mentions that the DeT is firstly pretrained on Pseudo LaSOT and Pseudo Coco, and then finetuned on DepthTrack. I would like to know if it makes any changes if we directly train on both three datasets?

Secondly, on the github it mentions that Using the default DiMP50 or ATOM pretrained checkpoints can reduce the training time. It seems that DiMP or ATOM are pretrained with larger RGB datasets (trackingnet, got10k, etc etc). I would like to know if these pretrained weights are adopted to initialize the model weight to produce the paper results? Or in the paper the network is only trained with Pseudo LaSOT, Pseudo Coco, and DepthTrack.

Finally, one question regarding the Table 2: Comparison of the original RGB trackers and their DeT variants. How are the RGB baseline trained? Only with RGB images from Pseudo LaSOT, Pseudo Coco, and DepthTrack? Are they only initialized with pretrained encoder (Imagenet)?

Sorry to bother you with all these questions... Looking forward to hearing from you. Thanks again