kitti_15 results reproduction D1 is not align with the open weights

[luyao-cv]

thanks for your good work, but i have downloaded the sceneflow weights, and trained the kitti_15 datasets using the kitti_ft.py as the command "python KITTI_ft.py --data_path /data/kitti_2015/training/ --gpu_id 4 --load_path SceneFlow.pth --no_cuda --batch_size 4" , however, the results is not good as the open weights. can you provide the command to get the results as good as the open weight? Thanks very much!

[SpadeLiu]

thanks for your good work, but i have downloaded the sceneflow weights, and trained the kitti_15 datasets using the kitti_ft.py as the command "python KITTI_ft.py --data_path /data/kitti_2015/training/ --gpu_id 4 --load_path SceneFlow.pth --no_cuda --batch_size 4" , however, the results is not good as the open weights. can you provide the command to get the results as good as the open weight? Thanks very much!

Hi,

It is better to know how bad your model performs. If very poor, maybe there are some bugs in the code, resulting in a failed training process.

If the 3px-Error evaluation results are around 2%, the training is successful. Under this condition, if you test the open weight and the result is much better (maybe around 1%), one of the reasons is that the weight we upload is trained with all of the training pairs (200 rather than 160), which is used to evaluate on the online KITTI benchmark. Another possible reason is that our evaluation sets are different, my evaluation set might cover your training set.

On the other hand, if you test the open weight and the result is just a little better (maybe around 1.8%), that might be caused by the different environments we use, and I suggest you retrain a model on SceneFlow and use it to finetune on KITTI.