4 The motion search is one of the two major components of DIS. It works more or less
5 like you'd expect; there's a bunch of overlapping patches (8x8 or 12x12 pixels) in
6 a grid, and for each patch, there's a search to try to find the most similar patch
9 Unlike in a typical video codec, the DIS patch search is based on gradient descent;
10 conceptually, you start with an initial guess (the value from the previous level,
11 or the zero flow for the very first level), subtract the reference (“template”)
12 patch from the candidate, look at the gradient to see in what direction there is
13 a lower difference, and then inch a bit toward that direction. (There is seemingly
14 nothing like AdaM, Momentum or similar, but the searched value is only in two
15 dimensions, so perhaps it doesn't matter as much then.)
17 DIS does a tweak to this concept. Since the procedure as outlined above requires
18 computing the gradient of the candidate patch, it uses the reference patch as
19 candidate (thus the “inverse” name), and thus uses _its_ gradient to understand
20 in which direction to move. (This is a bit dodgy, but not _that_ dodgy; after
21 all, the two patches are supposed to be quite similar, so their surroundings and
22 thus also gradients should also be quite similar.) It's not entirely clear whether
23 this is still a win on GPU, where calculations are much cheaper, especially
24 the way we parallelize the search, but we've kept it around for now.
26 The inverse search is explained and derived in the supplementary material of the
27 paper, section A. Do note that there's a typo; the text under equation 9 claims
28 that the matrix H is n x n (where presumably n is the patch size), while in reality,
31 Our GPU parallellization is fairly dumb right now; we do one patch per fragment
32 (ie., parallellize only over patches, not within each patch), which may not
33 be optimal. In particular, in the initial level, we only have 40 patches,
34 which is on the low side for a GPU, and the memory access patterns may also not
40 flat in int ref_layer, search_layer;
43 uniform sampler2DArray flow_tex, image_tex;
44 uniform usampler2DArray grad_tex; // Also contains the corresponding reference image.
45 uniform vec2 inv_image_size, inv_prev_level_size;
46 uniform uint patch_size;
47 uniform uint num_iterations;
49 vec3 unpack_gradients(uint v)
52 uint xi = (v >> 8) & 0xfffu;
54 vec3 r = vec3(xi * (1.0f / 4095.0f) - 0.5f, yi * (1.0f / 4095.0f) - 0.5f, vi * (1.0f / 255.0f));
58 // Note: The third variable is the actual pixel value.
59 vec3 get_gradients(vec3 tc)
61 vec3 grad = unpack_gradients(texture(grad_tex, tc).x);
63 // Zero gradients outside the image. (We'd do this with a sampler,
64 // but we want the repeat behavior for the actual texels, in the
66 if (any(lessThan(tc.xy, vec2(0.0f))) || any(greaterThan(tc.xy, vec2(1.0f)))) {
75 vec2 image_size = textureSize(grad_tex, 0).xy;
77 // Lock the patch center to an integer, so that we never get
78 // any bilinear artifacts for the gradient. (NOTE: This assumes an
79 // even patch size.) Then calculate the bottom-left texel of the patch.
80 vec2 base = (round(patch_center * image_size) - (0.5f * patch_size - 0.5f))
83 // First, precompute the pseudo-Hessian for the template patch.
84 // This is the part where we really save by the inverse search
85 // (ie., we can compute it up-front instead of anew for each
90 // where S is the gradient at each point in the patch. Note that
91 // this is an outer product, so we get a (symmetric) 2x2 matrix,
94 vec2 grad_sum = vec2(0.0f); // Used for patch normalization.
95 float template_sum = 0.0f;
96 for (uint y = 0; y < patch_size; ++y) {
97 for (uint x = 0; x < patch_size; ++x) {
98 vec2 tc = base + uvec2(x, y) * inv_image_size;
99 vec3 grad = get_gradients(vec3(tc, ref_layer));
100 H[0][0] += grad.x * grad.x;
101 H[1][1] += grad.y * grad.y;
102 H[0][1] += grad.x * grad.y;
104 template_sum += grad.z; // The actual template pixel value.
110 // Make sure we don't get a singular matrix even if e.g. the picture is
111 // all black. (The paper doesn't mention this, but the reference code
112 // does it, and it seems like a reasonable hack to avoid NaNs. With such
113 // a H, we'll go out-of-bounds pretty soon, though.)
114 if (determinant(H) < 1e-6) {
119 mat2 H_inv = inverse(H);
121 // Fetch the initial guess for the flow, and convert from the previous size to this one.
122 vec2 initial_u = texture(flow_tex, flow_tc).xy * (image_size * inv_prev_level_size);
124 float mean_diff, first_mean_diff;
126 for (uint i = 0; i < num_iterations; ++i) {
127 vec2 du = vec2(0.0, 0.0);
128 float warped_sum = 0.0f;
129 vec2 u_norm = u * inv_image_size; // In [0..1] coordinates instead of pixels.
130 for (uint y = 0; y < patch_size; ++y) {
131 for (uint x = 0; x < patch_size; ++x) {
132 vec2 tc = base + uvec2(x, y) * inv_image_size;
133 vec3 grad = get_gradients(vec3(tc, ref_layer));
135 float warped = texture(image_tex, vec3(tc + u_norm, search_layer)).x;
136 du += grad.xy * (warped - t);
137 warped_sum += warped;
141 // Subtract the mean for patch normalization. We've done our
142 // sums without subtracting the means (because we didn't know them
145 // sum(S^T * ((x + µ1) - (y + µ2))) = sum(S^T * (x - y)) + (µ1 – µ2) sum(S^T)
147 // which gives trivially
149 // sum(S^T * (x - y)) = [what we calculated] - (µ1 - µ2) sum(S^T)
151 // so we can just subtract away the mean difference here.
152 mean_diff = (warped_sum - template_sum) * (1.0 / float(patch_size * patch_size));
153 du -= grad_sum * mean_diff;
156 first_mean_diff = mean_diff;
159 // Do the actual update.
163 // Reject if we moved too far. Note that the paper says “too far” is the
164 // patch size, but the DIS code uses half of a patch size. The latter seems
165 // to give much better overall results.
167 // Also reject if the patch goes out-of-bounds (the paper does not mention this,
168 // but the code does, and it seems to be critical to avoid really bad behavior
170 vec2 patch_center = (base * image_size - 0.5f) + patch_size * 0.5f + u;
171 if (length(u - initial_u) > (patch_size * 0.5f) ||
172 patch_center.x < -(patch_size * 0.5f) ||
173 image_size.x - patch_center.x < -(patch_size * 0.5f) ||
174 patch_center.y < -(patch_size * 0.5f) ||
175 image_size.y - patch_center.y < -(patch_size * 0.5f)) {
177 mean_diff = first_mean_diff;
180 // NOTE: The mean patch diff will be for the second-to-last patch,
181 // not the true position of du. But hopefully, it will be very close.
183 out_flow = vec3(u.x, u.y, mean_diff);