jmtatsch

Is scipy's misc.imresize(img, self.input_shape) really exactly the same as matlabs imresize(img,[new_rows new_cols],'bilinear'); How about the tf.image.resize vs the caffe Interp layer?

Looking into this further, there are minor differences between the python and Matlab image resizing depending on input data type. See e.g. https://stackoverflow.com/questions/26812289/matlab-vs-c-vs-opencv-imresize Matlab performs anti-aliasing by default and also...

@hujh14 Unfortunately, i cannot run the original code on a 8GB 1080 GTX, not even with batch size 1 due to insufficient memory. Did you manage to compile the original...

@Vladkryvoruchko are you sure it works with the cityscapes model? it is much larger than both others... which CUDA, CuDNN?

Doing a sliced evaluation now, much better detail!  Will do a smarter overlap treatment then evaluate again

#13

Without flipped evaluation ``` classes IoU nIoU -------------------------------- road : 0.981 nan sidewalk : 0.849 nan building : 0.922 nan wall : 0.572 nan fence : 0.624 nan pole :...

Okay, with flipped evaluation: ``` classes IoU nIoU -------------------------------- road : 0.982 nan sidewalk : 0.853 nan building : 0.924 nan wall : 0.582 nan fence : 0.634 nan pole...

Adding multi-scale on top actually worsened the results by far. A funky gif with scales = [0.5, 0.75, 1.0, 1.25, 1.5, 1.75] last frame is the aggregated one. ![evaluation_of_scales_2017-09-05 10...

@leinxx by sliced prediction I mean cutting the image into 4x2 overlapping 713x713 slices, forwarding them though the network and reassembling the 2048x1024 predictions from them.  Please let us...