Gaussian counting heatmap

[Northern-mountains]

It's an excellent job. Here I'd like to ask how to draw the Gaussian counting heatmap in the paper. This is a very interesting graph. Can you open source the code for this part?

[tail-3lbs]

Hi,

Thank you for your interest in our work.

The key information is stored in gaussian_count_per_tile in trainer.py.

The following snippets are used to generate the Gaussain count heatmap.

from PIL import Image

def __save_heatmap(arr: np.ndarray, path: str, vmin: float, vmax: float, cmap: str = 'viridis'):
    """Save a 2D array as a heatmap image at exact pixel dimensions (no matplotlib figure margins)."""
    # Map fixed [vmin, vmax] range to [0, 1] as required by the colormap API.
    # This is NOT per-image normalization: vmin/vmax are fixed constants, so the
    # same value always maps to the same color across all images (equivalent to
    # matplotlib's ax.imshow(arr, vmin=vmin, vmax=vmax)).
    scaled = np.clip((arr - vmin) / max(vmax - vmin, 1e-8), 0, 1)
    colored = (plt.get_cmap(cmap)(scaled)[:, :, :3] * 255).astype(np.uint8)
    Image.fromarray(colored).save(path)


tiles_x = int(math.ceil(gt_image.shape[3] / float(pp.tile_size)))
tiles_y = int(math.ceil(gt_image.shape[2] / float(pp.tile_size)))

clamped_gaussian_count = torch.clamp(gaussian_count_per_tile, 0, 1600)
count_map_tiles = clamped_gaussian_count.unsqueeze(1).unsqueeze(-1).unsqueeze(-1).expand(
    -1, 1, -1, pp.tile_size, pp.tile_size
)  # [batch, 1, num_tiles, tile_size, tile_size]
count_map = utils.tiles2img_torch(count_map_tiles, tiles_x, tiles_y)[..., :gt_image.shape[2], :gt_image.shape[3]].contiguous()

...

count_map_np = count_map[batch_idx, 0].cpu().numpy()  # [H, W]
__save_heatmap(count_map_np, heatmap_path, vmin=0, vmax=1600)