Datasets:
File size: 9,124 Bytes
58d236b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 |
"""
CHIP Dataset Usage Example
This script demonstrates how to load and visualize data from the CHIP dataset,
including RGB images, depth maps, camera parameters, and 3D object models.
Requirements:
pip install datasets huggingface_hub numpy opencv-python open3d torch
"""
import json
import os
import gc
from typing import Tuple, Optional
import numpy as np
import cv2
import open3d as o3d
import torch
from datasets import load_dataset, get_dataset_infos
from huggingface_hub import snapshot_download
def lift_point_cloud(
depth: torch.Tensor,
camera_intrinsics: torch.Tensor,
xy_indices: Optional[Tuple[torch.Tensor, torch.Tensor]] = None
) -> torch.Tensor:
"""
Lift a depth image to a 3D point cloud using camera intrinsics.
Args:
depth: Depth image tensor of shape (H, W, C) where C >= 1.
If C > 1, channels 1+ are treated as features (e.g., RGB).
camera_intrinsics: Flattened camera intrinsic matrix [fx, 0, cx, 0, fy, cy, 0, 0, 1].
xy_indices: Optional tuple of (x_coords, y_coords) to lift only specific pixels.
Returns:
Point cloud tensor of shape (N, 3+F) where F is the number of feature channels.
First 3 columns are XYZ coordinates, remaining columns are features.
"""
H, W, num_channels = depth.shape
depth_values = depth[:, :, 0]
if xy_indices is not None:
x_coords, y_coords = xy_indices
x_coords = x_coords.to(depth_values.device).float()
y_coords = y_coords.to(depth_values.device).float()
z_coords = depth_values[y_coords.long(), x_coords.long()]
else:
# Create pixel coordinate grids
x_grid, y_grid = np.meshgrid(
np.arange(W, dtype=np.float32),
np.arange(H, dtype=np.float32),
indexing='xy'
)
x_coords = torch.from_numpy(x_grid).flatten().to(depth_values.device)
y_coords = torch.from_numpy(y_grid).flatten().to(depth_values.device)
z_coords = depth_values.flatten()
# Extract camera intrinsics
fx, fy = camera_intrinsics[0], camera_intrinsics[4]
cx, cy = camera_intrinsics[2], camera_intrinsics[5]
# Back-project to 3D coordinates
x_3d = (x_coords - cx) * z_coords / fx
y_3d = (y_coords - cy) * z_coords / fy
points_3d = torch.stack([x_3d, y_3d, z_coords], dim=1)
# Add additional features (e.g., RGB) if present
if num_channels > 1:
features = depth[y_coords.long(), x_coords.long(), 1:]
if xy_indices is None:
features = features.reshape(H * W, num_channels - 1)
points_3d = torch.cat([points_3d, features], dim=1)
return points_3d
def back_project_rgbd(
rgb: np.ndarray,
depth: np.ndarray,
camera_intrinsics: np.ndarray
) -> torch.Tensor:
"""
Back-project RGB-D image to a colored point cloud.
Args:
rgb: RGB image array of shape (H, W, 3).
depth: Depth map array of shape (H, W) with values in meters.
camera_intrinsics: Flattened 3x3 camera intrinsic matrix.
Returns:
Point cloud tensor of shape (N, 6) with XYZ and RGB columns.
"""
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Get valid depth pixel coordinates
valid_rows, valid_cols = np.where(depth > 0)
xy_indices = torch.tensor(
np.stack([valid_cols, valid_rows]),
dtype=torch.long,
device=device
)
# Concatenate depth and RGB channels
depth_rgb = torch.cat([
torch.from_numpy(depth).unsqueeze(-1),
torch.from_numpy(rgb)
], dim=2).to(device)
# Lift to 3D point cloud
camera_tensor = torch.from_numpy(camera_intrinsics).to(device)
point_cloud = lift_point_cloud(depth_rgb, camera_tensor, tuple(xy_indices))
return point_cloud
def visualize_chip_sample(
repo_id: str = "FBK-TeV/CHIP",
target_dir: str = "./chip_data",
num_samples: int = 1,
show_2d: bool = False
) -> None:
"""
Load and visualize samples from the CHIP dataset.
Args:
repo_id: Hugging Face dataset repository ID.
target_dir: Local directory to store downloaded model files.
num_samples: Number of samples to visualize.
show_2d: If True, display RGB and depth images in OpenCV windows.
If False, only show 3D point cloud visualization.
"""
# Display dataset information
info = get_dataset_infos(repo_id)
print(f"Dataset info: {info}\n")
# Download 3D object models
print("Downloading 3D models...")
local_path = snapshot_download(
repo_id=repo_id,
repo_type="dataset",
local_dir=target_dir,
allow_patterns=["models/*"]
)
print(f"Models downloaded to: {local_path}\n")
# Stream dataset samples
dataset = load_dataset(repo_id, streaming=True)
for idx, example in enumerate(dataset['test'].take(num_samples)):
print(f"Processing sample {idx + 1}/{num_samples}...")
# ========== Load RGB Image ==========
rgb_image = np.array(example['image'])
rgb_bgr = cv2.cvtColor(rgb_image, cv2.COLOR_RGB2BGR)
# ========== Load Depth Map ==========
depth_map = np.array(example['depth'], dtype=np.uint16).astype(np.float32)
# Visualize depth for display
depth_vis = cv2.normalize(depth_map, None, 0, 255, cv2.NORM_MINMAX)
depth_vis = depth_vis.astype(np.uint8)
# ========== Parse Camera Parameters ==========
camera_params = json.loads(example['camera_params'])
intrinsics_matrix = np.array(camera_params['cam_K']).reshape(3, 3)
depth_scale = camera_params['depth_scale']
print(f"Camera intrinsics:\n{intrinsics_matrix}")
print(f"Depth scale: {depth_scale}")
# ========== Parse Object Labels ==========
labels = json.loads(example['labels'])
label = labels[0] # Process first object
rotation_matrix = np.array(label['cam_R_m2c_flat']).reshape(3, 3)
translation_vector = np.array(label['cam_t_m2c'])
bbox = [
label['bbox_x'],
label['bbox_y'],
label['bbox_width'],
label['bbox_height']
]
print(f"\nObject ID: {label['obj_id']}")
print(f"Rotation matrix:\n{rotation_matrix}")
print(f"Translation vector: {translation_vector}")
print(f"Bounding box (x, y, w, h): {bbox}\n")
# ========== Visualize 2D ==========
if show_2d:
x, y, w, h = bbox
rgb_with_bbox = rgb_bgr.copy()
cv2.rectangle(rgb_with_bbox, (x, y), (x + w, y + h), (0, 255, 0), 2)
cv2.imshow("RGB Image", rgb_with_bbox)
cv2.imshow("Depth Map", depth_vis)
print("Displaying 2D images (press any key to continue to 3D)...")
cv2.waitKey(0)
# ========== Create 3D Point Cloud ==========
depth_metric = depth_map * depth_scale
point_cloud = back_project_rgbd(rgb_image, depth_metric, intrinsics_matrix.flatten())
# Convert to Open3D format
pcd_o3d = o3d.geometry.PointCloud()
pcd_o3d.points = o3d.utility.Vector3dVector(point_cloud[:, :3].cpu().numpy())
if point_cloud.shape[1] > 3:
pcd_o3d.colors = o3d.utility.Vector3dVector(
point_cloud[:, 3:].cpu().numpy() / 255.0
)
# ========== Load and Transform 3D Model ==========
model_path = os.path.join(target_dir, "models", f"obj_{label['obj_id']:06d}.ply")
model_mesh = o3d.io.read_triangle_mesh(model_path)
model_mesh.paint_uniform_color([1.0, 0.0, 0.0]) # Red
# Apply pose transformation
pose_matrix = np.eye(4)
pose_matrix[:3, :3] = rotation_matrix
pose_matrix[:3, 3] = translation_vector
model_mesh.transform(pose_matrix)
# ========== Visualize 3D ==========
print("Displaying 3D visualization (close window to continue)...")
o3d.visualization.draw_geometries(
[pcd_o3d, model_mesh],
window_name=f"CHIP Sample {idx + 1}: Scene + Model (Red)"
)
if show_2d:
cv2.destroyAllWindows()
else:
# Small delay to prevent visualization from closing too quickly
cv2.waitKey(100)
# Cleanup
del dataset
gc.collect()
print("\nVisualization complete!")
if __name__ == "__main__":
# Example usage
# Option 1: Show both 2D images and 3D point cloud
visualize_chip_sample(
repo_id="FBK-TeV/CHIP",
target_dir="./chip_data",
num_samples=1,
show_2d=True # Display RGB and depth images
)
# Option 2: Show only 3D point cloud visualization
# visualize_chip_sample(
# repo_id="FBK-TeV/CHIP",
# target_dir="./chip_data",
# num_samples=1,
# show_2d=False # Skip 2D visualization
# ) |