| 66 | self.wheel = self.wheel / 255. |
| 67 | |
| 68 | def visualize(self, nnf): |
| 69 | assert len(nnf.shape) == 3 |
| 70 | assert nnf.shape[2] == 2 |
| 71 | |
| 72 | RY, YG, GC = 15, 6, 4 |
| 73 | YG, GC, CB = 6, 4, 11 |
| 74 | BM, MR = 13, 6 |
| 75 | NCOLS = RY + YG + GC + CB + BM + MR |
| 76 | |
| 77 | fx = nnf[:, :, 0].astype(np.float32) |
| 78 | fy = nnf[:, :, 1].astype(np.float32) |
| 79 | |
| 80 | h, w = fx.shape[:2] |
| 81 | fx = fx.reshape([-1]) |
| 82 | fy = fy.reshape([-1]) |
| 83 | |
| 84 | rad = np.sqrt(fx * fx + fy * fy) |
| 85 | |
| 86 | normalizer = max(rad.max(), 1) |
| 87 | # This parameter controls how sensitive the visualization is to small displacement |
| 88 | # The smaller it is, the more sensitive the visualization is. |
| 89 | # We don't let it be smaller than 1 since we don't want to be sensitive to noise. |
| 90 | |
| 91 | a = np.arctan2(-fy, -fx) / np.pi |
| 92 | fk = (a + 1.0) / 2.0 * (NCOLS - 1) |
| 93 | k0 = fk.astype(np.int32) |
| 94 | k1 = (k0 + 1) % NCOLS |
| 95 | f = (fk - k0).astype(np.float32) |
| 96 | |
| 97 | color0 = self.wheel[k0, :] |
| 98 | color1 = self.wheel[k1, :] |
| 99 | |
| 100 | f = np.stack([f, f, f], axis=-1) |
| 101 | color = (1 - f) * color0 + f * color1 |
| 102 | |
| 103 | color = 1 - (np.expand_dims(rad, axis=-1) / normalizer) * (1 - color) |
| 104 | |
| 105 | return color.reshape(h, w, 3)[:, :, ::-1] |
| 106 | |
| 107 | |
| 108 | if __name__ == '__main__': |