zyayoung

我想知道真机的global odometry是怎么实现的？ 不知道作者是否能简单介绍下真机部署的流程，谢谢！

Localization Drift Compensation这部分的代码也是开源的吗 😢 还烦请作者指一下路，不胜感激

Also, mobile WIFI hotspots seem unstable, even at close distances. How did you tackle communication with contral terminals and multiple drones?

> are you submitting to our cvpr challenge server? Could you please tell us your challenge ID? Yes. the challenge ID is [1259](https://eval.ai/challenge/1259/).

We are not using the bdd100k image detection set, which can improve the detection performance. The reported results are mMOTA and mIDF1.

We are using the third config, _i.e._, `r50.bdd100k_mot.20e`. Sorry for causing confusion.

The total training time was 6d18h on 8 2080ti GPUs.

1. The pretrained model is for testing on the MOT17 test split. 2. We use pre-trained Deformable DETR **+ iterative bounding box refinement** from [Deformable-DETR](https://github.com/fundamentalvision/Deformable-DETR#main-results). You may try this [weight](https://drive.google.com/file/d/1JYKyRYzUH7uo9eVfDaVCiaIGZb5YTCuI/view?usp=sharing)....

Thank you so much for your interest in our work! For GPU memory limitation, we now support training with 11GB memory or less by adding `--use_checkpoint` argument. To save more...

Have you tried data preparation to generate txt files in `labels_with_ids`? Maybe you could try data preparation in [FairMOT](https://github.com/ifzhang/FairMOT).