what is the performance on kitti train

[LiewFeng]

Hi, @Cc-Hy , I wonder the performance on the kitti train. Can you kindly provide it? Lots of thanks!

[Cc-Hy]

@LiewFeng Here are two typical results trained with KITTI train only.

And you can also refer to our paper for more results.

[LiewFeng]

Thank you very much for your quick reply! I run the code for 2 times and get a slightly different results from the paper. And as you mentioned in the README, this implementation has some differences from your paper. That's the reason I wonder the performance on the kitti train. The performance on hard is about 1 point lower than taht on the paper and the performance on easy is 1 point higher than that on the paper but not very stable. The results is obtained with 1 GPU. I will try 2 GPUs latter. By the way, does it affect performance much?

[Cc-Hy]

@LiewFeng Hi, I believe it is a common seen situation that the results trained with KITTI train is unstable. For one thing, the number of KITTI train samples is small and the performance often varies 1-2 point or more, and this happens a lot when I'm training my code or using someone else's repo. For another, different softwares and hardwares may also cause a difference.

So if you can not exactly reproduce our results, just set your own results as the baseline and do the experiments. Also, BEV based/dense depth based methods need more training samples to be powerful and stable, you may use the unlabeled samples for a try.

[LiewFeng]

Thanks. I will try the unlabeled smaples latter. The downloading of raw data really takes time.

[Cc-Hy]

@LiewFeng I think so, too. And the data of Waymo and Nuscenes is even bigger, so take your time.

[LiewFeng]

Okay. Looking forward to your code on these two datasets.