Reconstruction accuracy violations for negatives

[qxcv]

The check on line 288 of cy/src/detect.m, which checks that the score of a pose found through message passing matches that found through reconstruction of the pose features from type and location information, keeps triggering. This check was inherited from Yang & Ramanan's original code, although I'm pretty sure it should still be valid for my mode. I've reproduced the check below:

if wrote_ex
    w = -(qp.w + qp.w0.*qp.wreg) / qp.Cneg;
    sv_score = score(w,qp.x,qp.n);
    sv_rscore = rscore(y,x,t);
    delta = abs(sv_score - sv_rscore);
    if delta >= 1e-5
        fprintf(...
            'Delta %f = |%f (SV score) - %f (rscore)| too big\n', ...
            delta, sv_score, sv_rscore);
    end
end

The deltas reported are quite significant (i.e. of the same magnitude as the two scores).

[qxcv]

Possible causes I can think of (will update list as I go):

• [x] I may not be scaling displacement vectors correctly during Gaussian-of-deformation reconstruction. Alternately, I may not be scaling them correctly when calculating them in the first place (in passmsg). There are two types of down/up-sampling to be aware of when scaling: the explicit resampling I do to get multi-scale detections, and the implicit downsampling the CNN performs due to pooling and non-unity strides.
• [x] Uh, could the definition of square as int -> int in shiftdt.cpp have something to do with it? Especially given that I'm not actually passing it an int?

[qxcv]

Yup, it was probably the int square(int); problem. No assertions trigger after changing that to template<class T> T square(T); (see 2e1311b9c596244e83c8fb9ac5dfd9fbe191a4b3).