yahooarchive / crow Goto Github PK

Hi~When I was using 'python extract_features.py --images oxford/data/* --out oxford/pool5' It is slow so I modify the code extract_features.py by adding 'caffe.set_mode_gpu()' in the line 73, but it immediately throws out an error: Check failed : error == cudaSuccess (2 vs. 0) out of memory
I am using Tesla K20 with 4GB of GPU memory. Can someone help me proceed from here?

I think line 86 in crow.py should be fixed by:
return sknormalize(x.reshape(1,-1), copy=copy)

Hi, i do the same experiments all mentioned in your paper. However,i can not get the same results, i just can not understand where am i wrong. I just wonder did somebody get the same results? Many thanks!

Hi -- how does crow deal with resizing images? I was under the impression that VGG16 took a 224x224 image as input, but it doesn't look like anything's being resized in the code.

Thanks
Ben

if we crop the image as the query, you define the box as
box = map(lambda d: int(round(d)), (x, y, x + w, y + h))
however, the four numbers in the gt file represents the x1,y1,x2,y2 respectively, rather than x1,y1,w,h
so the cropped image is not right.
and the map of the wrongly cropped query image is better than the correct version

Hi, I'd like to raise a question. Here is the question.
When I do some experiments on the Holiday dataset, I can not get the same results. There are almost 20 percents gaps between my results and yours. So I wonder if I can have one copy of your source code which contains the Holiday dataset part.
Many thanks in advance.

“The version here keeps the full resolution for the best possible retrieval performance” in paper， but the input size is 10 * 3 * 224 * 224 in VGG_ILSVRC_16_pool5.prototxt. Will it resize the img into 224 * 224 automatically when running?
The problem I met is the poor performance when applying to imgs whose "height > width".
@pumpikano

Should they have the same distribution with the under test data @yahoocla