GameIR / demo /utils.py
LLLebin's picture
Add demo
9446d15 verified
raw
history blame
4.84 kB
import numpy as np
def get_depth_map(depth_map_raw: np.ndarray) -> np.ndarray:
"""
Converts a raw RGB-encoded depth map to actual one channel depth map.
Parameters:
depth_map_raw (np.ndarray): The raw RGB depth map where each pixel's RGB values encode depth information.
The shape should be (height, width, 3).
Returns:
np.ndarray: The converted depth map where each pixel value represents the depth at that point in [0, 1000] meters.
The shape of the returned array is the same as the input but with only one channel.
Description:
The depth map camera captures a scene by encoding the distance of each pixel to the camera (also known as the depth buffer or z-buffer) into a 24-bit floating point precision image, codified across the RGB color space channels in the order of R --> G -> B .
R G B int24
00000000 00000000 00000000 0 min (near)
11111111 11111111 11111111 16777215 max (far)
The depth map is calculated by converting the RGB values to a single floating-point number using the formula:
depth = (R + G * 256 + B * 256^2) / (256^3 - 1) * 1000
Here R, G, B are the red, green, and blue channel values respectively, and the formula maps these to a range from 0 to 1000.
For more details, please refer to: https://carla.readthedocs.io/en/latest/ref_sensors/#depth-camera
"""
R = depth_map_raw[..., 0].astype(np.float32)
G = depth_map_raw[..., 1].astype(np.float32)
B = depth_map_raw[..., 2].astype(np.float32)
depth_map = (R + G * 256 + B * 256 * 256) / (256 * 256 * 256 - 1) * 1000
return depth_map
def get_segmentation_map(segmentation_map_raw: np.ndarray, colorize=False) -> np.ndarray:
"""
Extracts a segmentation map from a raw color segmentation image. Depending on the 'colorize' flag,
this function either returns a single-channel map or a colorized segmentation map where each label
is mapped to a specific RGB color defined by the label_colors dictionary.
Parameters:
segmentation_map_raw (np.ndarray): The raw color segmentation image, with the shape (height, width, 3).
Typically, the first channel (R) is used to represent segmentation information.
colorize (bool): If True, returns a colorized segmentation map. If False, returns the original
segmentation map's R channel as a float32 array.
Returns:
np.ndarray: If colorize is False, returns the extracted single-channel segmentation map,
where values are floating-point, with the shape (height, width). If colorize is True,
returns a 3-channel RGB image where each segmentation label is mapped to a predefined color.
Description:
The semantic segmentation camera classifies every object in the view by displaying it in a different color
according to the object class. For example, pedestrians appear in a different color than vehicles.
It provides an image with the tag information encoded in the red channel. A pixel with a red value of x
displays an object with tag x. When 'colorize' is True, each pixel's label is converted to a specific color
based on a predefined dictionary mapping labels to colors.
For more details, please refer to: https://carla.readthedocs.io/en/latest/ref_sensors/#semantic-segmentation-camera
"""
if colorize:
label_colors = {
0: (0, 0, 0),
1: (128, 64, 128),
2: (244, 35, 232),
3: (70, 70, 70),
4: (102, 102, 156),
5: (190, 153, 153),
6: (153, 153, 153),
7: (250, 170, 30),
8: (220, 220, 0),
9: (107, 142, 35),
10: (152, 251, 152),
11: (70, 130, 180),
12: (220, 20, 60),
13: (255, 0, 0),
14: (0, 0, 142),
15: (0, 0, 70),
16: (0, 60, 100),
17: (0, 60, 100),
18: (0, 0, 230),
19: (119, 11, 32),
20: (110, 190, 160),
21: (170, 120, 50),
22: (55, 90, 80),
23: (45, 60, 150),
24: (157, 234, 50),
25: (81, 0, 81),
26: (150, 100, 100),
27: (230, 150, 140),
28: (180, 165, 180)
}
height, width = segmentation_map_raw.shape[:2]
rgb_image = np.zeros((height, width, 3), dtype=np.uint8)
for label, color in label_colors.items():
rgb_image[segmentation_map_raw[..., 0] == label] = color
return rgb_image
else:
return segmentation_map_raw[..., 0].astype(np.float32)