crfasrnn
crfasrnn copied to clipboard
torrvision
crf-rn demo vs trained algorithm
I've processed the same image through the demo (http://www.robots.ox.ac.uk/~szheng/crfasrnndemo/classify_upload) and the installed trained classifier (TVG_CRFRNN_COCO_VOC.prototxt), and I get very different results (the ones on he website are way better). An suggestion why this is the case?
We used the same model running on the background for the demo. This could be caused by the scaling issue. If you have an input image with significant large size, demo website will resize them and then feed into the model.
I don't think it is very large, maybe 950x500. Any other possible reason you think?
If it is a gray input image, the demo script will convert them into 3-channels like image, each channel will be filled with the gray image.
For an image with resolution 950x500, you could try to resize the image into 475x250 (bilinear interpretation and keeping the ratio) and then try the demo script to see if it works the same.
OK, so I use the first image 1502x640 in the script and in the online demo. The result from the demo is far better. When I do it with a color image, overall results are worse, but the gap between demo and script is narrower.
Hi I don't mean to spam this board with complaining, but I am committed to making this repo work for me. I'm seeing similarly terrible results with the sample model, in CPU and GPU mode. Since my runtest fails at the multi stage meanfield test, maybe there's something just wrong with my code? By fail I mean the gradients are always dramatically incorrect for CPU, and the test seemingly gets stuck in GPU mode ( #103 )
@nathanin Using an old version of Caffe will have the same result of the demo. I can confirm that. I also met same issue as you, running demo and got a lame results. I think that's because we use a newer version of caffe and need to set the crop offset. I am quite frustrated by the versioning of this repo but by searching around, eventually, still making both training and testing work.
