dvl-tum / motsynth-baselines Goto Github PK

Thanks for sharing MOTSynth dataset and the corresponding code!

I have one minor comment for the paper. On paper P7 section 4.4. Multi-object Tracking left-column Results paragraph, the paper says "we obtain 45.0 MOTA and 51.2 IDF1 with our MOTSynth trained model, yielding +3.5 MOTA and +1.6 IDF1 improvement over the COCO trained model (43.5 MOTA and 49.6 IDF1).". Does it seem the improvement on MOTA should be +1.5 MOTA?

Thank you!

Hi, I found that some bounding box annotations are labeled as area=0 but still appear in the gt.txt file.

For example, in the 0089 video, frame 21:

{
"id": 890020000466,
"image_id": 890020,
"category_id": 1,
"segmentation":
{
"size":
[
1080,
1920
],
"counts": "PPYo1"
},
"area": 0.0,
"bbox":
[
854,
562,
4,
27
]
}

I am wondering do you have a threshold to determine if the area should be set as 0.0? Shall we modify this area value or remove this object from the gt.txt?

Thank you!

Hi author of MOTSYNTH
Can we use real data (e.ge. MOT17) with un-labeled and use unsupervised learning?

In the annotation file, I noticed that there are two fields named "cam_world_pos" and "cam_world_rot", what are the typical use cases of these two parameters? If there are any examples or explanations, that would be greatly appreciated. Thank you!

Why to frames need to substract 3?
I visulize the image finding the annotations is inconsistent with the generated images. In addition, when I remove the code of 'substract 3', annotations becom consistent.

Hasn't it been revealed yet?

Hi, Thank you for opening great dataset.

When MOTSynth is downloaded, sequences 629 and 757 are not provided in the training dataset, but are included in the train.txt split.
There are 574 sequences in total.
Will they be updated?

I encountered something strange while visualizing the dataset. As shown below, seq 012, 0110.jpg:


The text upon the bbox represents the identity and visibility.

The 65 and 38 identities seem to be annotated incorrectly. I think MOTSynth is automatically annotated through some code. I wonder to know what causes this incorrectness.

BTW, I also want to ask about the meaning of visibility in MOT annotation. Is it calculated by the bounding box area or the mask area?

THANKS A LOT.

the names of downloaded reid images don't follow the normal rules. I don't know the real id and camera id of a certain image. Thanks for your reply.

where is environment.yml file?

Could you please provide more details regarding the YOLOv3 baseline?

Since most of the common configurations (e.g., YOLOv3-tiny, YOLOv3, YOLOv3-SPP, YOLOv3-SPP-ultralytics) all accept the input size of 608.

Thank you in advance!

Hi, I have two questions about this MOTSynth challenge

1. After following data preparing process as following site mot annotations has starting from repo '000'.
   
   However, extracted frames from original zip file is starting from '003', so what am I missing in this stage? It looks like wget code suppose to get zip file from sequence including '000'.
   

https://github.com/dvl-tum/motsynth-baselines/blob/main/docs/DATA_PREPARATION.md

2. According to evaluation server, it suppose to add not only MOT17-test.txt but also MOTSynth.txt. Can't we just submit MOT17-test.txt for saving time? (because evaluation also takes one-hour or more for only MOT17)
   

Thanks

Hi.
I want to try to integrate depth information to train my multi-object tracker, but I didn't find them in the downloaded MOTSynth dataset. When will you upload the depth information of the dataset?
Looking forward to your reply.