Regarding reference for performance of GCN "(pixels baseline) [19], low res" provided in table 2 for this work.

[aryan-at-ul]

Hi,
I wanted to reach out to you regarding the accuracy of GCN (pixels baseline)[19] table 2, I was working on something very similar, but I needed some previous work to compare with, my performance is very poor not more than 34% when I checked the given reference it has no mention of GCN being applied for Cifar10 classification. Was that the result you gained when converting data in this format
data = [torch.from_numpy(np.concatenate((coord, avg_values), axis=1)).unsqueeze(0).float(), torch.from_numpy(A_spatial).unsqueeze(0).float(), False]
where the number of superpixels is equal to the total pixels present?
Any insight you can give me would be constructive. I am looking to refer to the application of GCN for cifar10 classification, where nodes in the graph represent pixels in the image.

regards
Arayn

[bknyaz]

Hi, do you get 34% without coordinate features? The GCN [19] pixels baseline in my Table 2 achieving 50.57% uses both pixel values and coordinates so each node has 5 features (R,G,B,x,y). See paragraph "Graph formation for images" describing that. Also, in this baseline the images are downsampled to 12x12, so each graph is a regular 2d grid with144 nodes.
The paper "Benchmarking Graph Neural Networks" also has more experiments (with strong results on superpixels) and their github implementation is available.

[aryan-at-ul]

Thanks for the prompt response, I will add 2d grid i,j coordinate and then try, currently i experimented by adding x,y as ( i/w - 0.5, i/h - 0.5 ) based on positional augmentation in some previous work. Thank you will also try downsamping the image for baseline.

Thanks again.