search

MCG-NJU / MeMOTR

[ICCV 2023] MeMOTR: Long-Term Memory-Augmented Transformer for Multi-Object Tracking
https://arxiv.org/abs/2307.15700
MIT License
157 stars 11 forks source link

Training instructions for MOT 17 #10

Closed sawhney-medha closed 9 months ago

sawhney-medha commented 9 months ago

Hello! Can you please upload instructions to use the the training script for MOT17? Thank you!!

HELLORPG commented 9 months ago

My bad. I forgot to upload this config file into configs but only put it here.

Now, as this commit, I have already uploaded the config file. Thanks a lot for your reminder.

sawhney-medha commented 9 months ago

Thank you so much!!

HELLORPG commented 9 months ago

I think you may be conducting experiments on MOT17, and there are some things that I need to explain to you. I did not reply earlier because I was too busy in the past several days. Sorry about that.

First of all, due to the merging of several versions of code, there was a problem in passing parameters during data processing, which I have corrected in this commit.

Moreover, there are some differences in the configuration on MOT17 (compared with DanceTrack/SportsMOT):

  1. The optimizer setting: 
    LR_POINTS: 2.0e-5    # 1.0e-5 on DanceTrack
WEIGHT_DECAY: 0.0001    # 0.0005 on DanceTrack

    I set the lr_points and weight_decay to 2.0e-5 and 0.0005 because of the stable convergence. This setting requires more training epochs, e.g. 18 (in our paper) to 20 (in our open-source code) on DanceTrack. However, as I discussed in #11, training on MOT17 is a delicate balance, so I haven't tried this new training configuration yet to see if it can achieve the desired results.

  2. The image size during augmentation: 
    COCO_SIZE: True    # False on DanceTrack

    This setting comes from MOTR. I don't quite understand the reason for this, but it does achieve better results.

  3. Whether the bounding boxes can exceed the image size: 
    OVERFLOW_BBOX: True    # False on DanceTrack

    Referring to some previous work, this approach can achieve better detection performance, but it will have a loss on the AssA metric. This is a difficult balance to choose.

BTW, it is also very common in MOT17 that newborn objects are suppressed by tracked targets, resulting in detection failure. Several recent research (CO-MOT/MOTRv3) focus on this. We hope we can resolve this issue in the future.