models/preresnet.py question

[sydney0zq]

models/preresnet.py question

[bearpaw]

@sydney0zq Yes you're right. The first conv should also be counted. Thus the total number of layers should be 9*3*3 + 3 = 84.

BTW, preresnet.py was just used for debugging at the beginning. It has never been evaluated for human pose estimation. You should try the hourglass model instead.

[sydney0zq]

@bearpaw
Thanks for your check.
And another question in models/hourglass.py. It may be not an issue, the origin code in Lua puzzles me a lot.

I made several trials on this hourglass. Hourglass._make_hour_glass actually construct a model which may be like:

When I set depth=1, one time of Upsampling and Downsampling operation will be forwarded. So it means that depth controls the Upsampling and Downsampling times?

I am just a novice of pose-estimation, so thanks a lot.

[bearpaw]

Yes. The 64x64 feature maps will be down sampled by 2^depth times and then upsampled to the original resolution (64x64).

[sydney0zq]

Hey, thanks for your patience.

But, hmmmmmm, I still have some points that I could not understand. Weeeeell, one more question
in eval_PCKh.m

% 0  - r ankle, 1 - r knee, 2 - r hip, 3 - l hip, 4 - l knee, 
% 5  - l ankle, 6 - pelvis, 7 - thorax, 8 - upper neck, 9 - head top, 
% 10 - r wrist, 11 - r elbow, 12 - r shoulder, 13 - l shoulder, 
% 14 - l elbow, 15 - l wrist
pa = [2, 3, 7, 7, 4, 5, 8, 9, 10, 0, 12, 13, 8, 8, 14, 15];

in showskeletons_joints.m:

            for child = 1:p_no
                if pa(child) == 0  % removed pa == 1, I think unnecessary 
                    continue;
                end
                x1 = x(pa(child));
                y1 = y(pa(child));
                x2 = x(child);
                y2 = y(child);
                
                plot(x1, y1, 'o', 'color', partcolor{child}, ... 
                    'MarkerSize', msize, 'MarkerFaceColor', partcolor{child});
                plot(x2, y2, 'o', 'color', partcolor{child}, ...
                    'MarkerSize', msize, 'MarkerFaceColor', partcolor{child});
                line([x1 x2], [y1, y2], 'color', partcolor{child}, 'linewidth', round(msize/2));
            end

What pa corrsponds for?

e.g. When we use pa(i), we take the pair joint of i? If it is, then upper neck may have two pairs, one for the head and the other for the thorax. Could you explain some mechanism about it?

Thanks again.

[bearpaw]

pa means the parent node of the current node. It determines which two body joints should be linked.