This version of code

[munchmoo]

HELLO,

I have a question regarding this repo.

Without BEV DA module and BEV KD loss, the student model is same as CaDDN. Does this repo version(using Resnet50 and some layer modification in detection, loss weight) of CaDDN produces comparable results with CaDDN(Resnet101) ?

+) in order to just reproduce the student model without KD loss, what config do I have to use?

Thank you.

[Cc-Hy]

Hi,

1. In fact, when I am using the official code of CaDDN (res101), I can not reproduce its performance on my device. And the results I get are 1-3 point lower than the given one. And the setting here (res50 and something else) has a similar performance. Something like this:

   
   Car AP_R40@0.70, 0.70, 0.70:
   bbox AP:97.8666, 89.5661, 84.1687
   bev  AP:27.9056, 20.8787, 18.4701
   3d   AP:20.5984, 14.7481, 12.5899
   aos  AP:97.06, 88.39, 82.54

Pedestrian AP_R40@0.50, 0.50, 0.50: bbox AP:52.0106, 44.0633, 38.3137 bev AP:15.8919, 11.2907, 8.9604 3d AP:12.5201, 8.7893, 6.9510 aos AP:25.18, 22.10, 19.52

Cyclist AP_R40@0.50, 0.50, 0.50: bbox AP:44.9809, 26.3292, 25.6032 bev AP:2.9665, 1.6742, 1.3987 3d AP:2.3066, 1.1249, 0.9944 aos AP:33.45, 19.37, 18.75

Not sure if it is caused by the hardware and software differences. So I recommend you to run it on your device and set it as the baseline.

2. To just use the student model,  you can use [this file](https://drive.google.com/file/d/1KAY3FzqeSaiP3uVUzdtJ6Whcq_YnS90o/view?usp=share_link). 

[munchmoo]

@Cc-Hy Is this result(res50 and something else)trained on kitti-train and evaluated on kitti-val?

[Cc-Hy]

Yes, and when we use the official settings, we also get similar results.

[Cc-Hy]

And the results are not very stable, but this is a typical one.

[munchmoo]

@Cc-Hy Thank you, I am trying to reproduce the Table 1 paper results. +) for the ablation study Does 'Pre' imply using depth-backbone pre-trained on KITTI-train for 40 epochs? Is this depth pretrained-backbone different from pre-trained DeeplabV3 model ? (which is provided in this repo)

[Cc-Hy]

Hi, to reproduce the Table 1 results, you should use the whole unlabeled data from kitti raw, and I will upload the split in a few days.

'Pre' means depth pre-training on kitti train, and the pre-trained DeeplabV3 model is pre-trained on COCO with 2D detection task, which is provided by torchvision, not depth pre-training.

[TimGor1997]

作者您好！请问如何在kitti上进行深度预训练呢？是否可以提供一下您使用的预训练checkpoint呢？我的邮箱是：595603009@qq.com，期待您的回复，谢谢！