evaluate problem

[D0ub1ePieR]

Is the evaluation progress like
input 1920 * 1080 alpha 1920 * 1080 trimap 1920 *1080
then resize input and trimap to 320 * 320
input 320 * 320 trimap 320 * 320 to the model, then get the prediction 320 * 320
prediction 320 * 320, resize to 1920 * 1080
pred 1920 * 1080 alpha 1920 * 1080
finally compare the resized prediction and the alpha(gt)

---

i test on the DIM-tensorflow version and cannot get the number in paper, because the sad,mse,conn and grad are effected by the image size
Thanks

[foamliu]

Please take a look at this script -- eval.py, you don't need to resize, the network can be adapted to different sizes.