| import torch |
| import numpy as np |
| import math |
| from trimesh.remesh import subdivide |
|
|
| def gaussian2D(shape, sigma=1): |
| m, n = [(ss - 1.) / 2. for ss in shape] |
| y, x = np.ogrid[-m:m+1,-n:n+1] |
|
|
| h = np.exp(-(x * x + y * y) / (2 * sigma * sigma)) |
| h[h < np.finfo(h.dtype).eps * h.max()] = 0 |
| return h |
|
|
| def calc_radius(bboxes_hw_norm, map_size=64): |
| if len(bboxes_hw_norm) == 0: |
| return [] |
| minimum_radius = map_size / 32. |
| scale_factor = map_size / 16. |
| scales = np.linalg.norm(np.array(bboxes_hw_norm)/2, ord=2, axis=1) |
| |
| radius = (scales * scale_factor + minimum_radius).astype(np.uint8) |
| return radius |
|
|
| def build_z_map(depth, with_coords = True, device='cpu'): |
| size = 2**depth |
| z_map = torch.zeros((size, size), dtype=torch.int64) |
| coords = torch.meshgrid(torch.arange(size),torch.arange(size),indexing='ij') |
| ys = coords[0] |
| xs = coords[1] |
| for i in range(depth): |
| z_map |= (xs & (1 << i)) << i | (ys & (1 << i)) << (i + 1) |
| if with_coords: |
| return z_map, ys.clone(), xs.clone() |
| else: |
| return z_map |
|
|
|
|
| def gen_scale_map(scales, v3ds, faces, cam_intrinsics, map_size, patch_size=28, pad=True): |
| if pad: |
| map_h = math.ceil(map_size[0]/2)*2 |
| map_w = math.ceil(map_size[1]/2)*2 |
| else: |
| map_h = map_size[0] |
| map_w = map_size[1] |
| scale_map = torch.zeros((map_h, map_w, 2)) |
|
|
| new_v3ds = [] |
| for v in v3ds: |
| vv, _ = subdivide(v, faces) |
| new_v3ds.append(torch.from_numpy(vv)) |
|
|
| v3ds = torch.stack(new_v3ds) |
|
|
| v2ds_homo = torch.matmul(v3ds,cam_intrinsics.transpose(-1,-2)) |
| v2ds = v2ds_homo[...,:2]/(v2ds_homo[...,2,None]) |
| v2ds_patch = (v2ds//patch_size).int() |
| v2ds_patch[..., 0] = v2ds_patch[..., 0].clamp(min = 0, max = map_size[1]-1) |
| v2ds_patch[..., 1] = v2ds_patch[..., 1].clamp(min = 0, max = map_size[0]-1) |
| for (v, s) in zip(v2ds_patch, scales): |
| scale_map[v[:, 1], v[:, 0], 0] = 1. |
| scale_map[v[:, 1], v[:, 0], 1] = s |
|
|
| return scale_map |
|
|
|
|
|
|
|
|
