gangweiX/ACVNet

About the pretrained model (pretrained_model/pretrained_model_sceneflow.ckpt) EPE:1.093, different from the result in Table 5

Closed this issue · 4 comments

A wonderful work!
I'd like to know whether the pretrained model corresponds to Table 5, because I tested the pretrained model in the test dataset of SceneFlow and got the result EPE:1.093, further larger than the result in Table 5.

Table 5 is results on KITTI, not Scene Flow. The pretrained model is for Scene Flow

Are the pretrained model the best? I ran the pretrained model you provided in the test dataset of SceneFlow, but just got EPE 1.093, further larger than the result in your paper.

You can use "python test_sceneflow.py" to test model on Scene Flow.

Thanks for replying. I got the same result as described in your paper this time. The previous result is caused by the compression of image files. To save storage, I compressed the original image files from the format 'PNG' to 'JPG'. Although there is little perceptual difference between the original and compressed, the final result diverges.