Datasets:
Tasks:
Image-to-3D
Modalities:
Geospatial
Languages:
English
Tags:
gaussian-splatting
novel-view-synthesis
3d-reconstruction
semantic-segmentation
remote-sensing
drone-imagery
License:
| """ | |
| EXAMPLE ONLY. Renders a trained 3D Gaussian Splatting model at this | |
| challenge's test camera poses, producing images that go straight into a | |
| submission's rgb/ folder. | |
| This script targets the vanilla method from | |
| https://github.com/graphdeco-inria/gaussian-splatting: it must be run from | |
| inside a gaussian-splatting (or API-compatible fork) checkout, next to | |
| train.py/render.py, since it imports that repo's | |
| scene/gaussian_renderer/utils/arguments modules. If your submission was | |
| produced with a different method/codebase, this script will not run as-is -- | |
| adapt the model loading and rendering calls to your own repo, keeping the | |
| same camera-pose parsing (cameras.txt/images.txt) and output naming | |
| (<output_dir>/rgb/<frame_id>.png) shown below. | |
| Output: | |
| <output_dir>/rgb/<frame_id>.png | |
| <frame_id> is the filename stem taken from the pose file's own image names | |
| (images.txt NAME column) -- the same stem the scoring program matches | |
| submitted renderings against. | |
| Usage Example (scene_000): | |
| python render_test_poses.py \\ | |
| --model_ply <model_path>/point_cloud/iteration_30000/point_cloud.ply \\ | |
| --camera_pose_dir Data_TUM/scene_000/test/sparse/0 \\ | |
| --output_dir <submission_dir>/scene_000 | |
| """ | |
| import torch | |
| import torchvision | |
| from pathlib import Path | |
| from argparse import ArgumentParser | |
| from scene.colmap_loader import read_extrinsics_text, read_intrinsics_text, qvec2rotmat | |
| from scene.gaussian_model import GaussianModel | |
| from scene.cameras import MiniCam | |
| from gaussian_renderer import render | |
| from utils.graphics_utils import getWorld2View2, getProjectionMatrix, focal2fov | |
| from arguments import PipelineParams | |
| def build_minicam(qvec, tvec, width, height, fx, fy, resolution_scale, znear=0.01, zfar=100.0): | |
| R = qvec2rotmat(qvec).transpose() | |
| T = tvec | |
| out_w = max(1, round(width / resolution_scale)) | |
| out_h = max(1, round(height / resolution_scale)) | |
| FoVx = focal2fov(fx, width) | |
| FoVy = focal2fov(fy, height) | |
| world_view_transform = torch.tensor(getWorld2View2(R, T)).transpose(0, 1).cuda() | |
| projection_matrix = getProjectionMatrix(znear=znear, zfar=zfar, fovX=FoVx, fovY=FoVy).transpose(0, 1).cuda() | |
| full_proj_transform = world_view_transform.unsqueeze(0).bmm(projection_matrix.unsqueeze(0)).squeeze(0) | |
| return MiniCam(out_w, out_h, FoVy, FoVx, znear, zfar, world_view_transform, full_proj_transform) | |
| def load_test_poses(camera_pose_dir: Path): | |
| cameras = read_intrinsics_text(camera_pose_dir / "cameras.txt") | |
| images = read_extrinsics_text(camera_pose_dir / "images.txt") | |
| entries = [] | |
| for image in images.values(): | |
| cam = cameras[image.camera_id] | |
| assert cam.model == "PINHOLE", "only undistorted PINHOLE cameras are supported" | |
| fx, fy = cam.params[0], cam.params[1] | |
| entries.append((image.name, image.qvec, image.tvec, cam.width, cam.height, fx, fy)) | |
| entries.sort(key=lambda e: e[0]) | |
| return entries | |
| def main(): | |
| parser = ArgumentParser(description="Render a trained 3DGS model at the challenge's test camera poses") | |
| parser.add_argument("--model_ply", required=True, type=Path, | |
| help="trained point_cloud.ply, e.g. <model_path>/point_cloud/iteration_30000/point_cloud.ply") | |
| parser.add_argument("--camera_pose_dir", required=True, type=Path, | |
| help="folder containing cameras.txt and images.txt for the test poses, " | |
| "e.g. Data_TUM/scene_000/test/sparse/0") | |
| parser.add_argument("--output_dir", required=True, type=Path, | |
| help="renders are written to <output_dir>/rgb/<frame_id>.png") | |
| parser.add_argument("--sh_degree", type=int, default=3) | |
| parser.add_argument("--resolution_scale", type=float, default=1.0, | |
| help="downsample factor applied to the cameras.txt WIDTH/HEIGHT; must match " | |
| "the reference image size the submission is scored against") | |
| parser.add_argument("--white_background", action="store_true") | |
| pipeline_params = PipelineParams(parser) | |
| args = parser.parse_args() | |
| pipeline = pipeline_params.extract(args) | |
| gaussians = GaussianModel(args.sh_degree) | |
| gaussians.load_ply(str(args.model_ply)) | |
| bg_color = [1, 1, 1] if args.white_background else [0, 0, 0] | |
| background = torch.tensor(bg_color, dtype=torch.float32, device="cuda") | |
| rgb_dir = args.output_dir / "rgb" | |
| rgb_dir.mkdir(parents=True, exist_ok=True) | |
| poses = load_test_poses(args.camera_pose_dir) | |
| print(f"Rendering {len(poses)} test views from {args.camera_pose_dir}") | |
| with torch.no_grad(): | |
| for name, qvec, tvec, width, height, fx, fy in poses: | |
| cam = build_minicam(qvec, tvec, width, height, fx, fy, args.resolution_scale) | |
| image = torch.clamp(render(cam, gaussians, pipeline, background)["render"], 0.0, 1.0) | |
| stem = Path(name).stem | |
| torchvision.utils.save_image(image, rgb_dir / f"{stem}.png") | |
| print(f"Done. renders -> {rgb_dir}") | |
| if __name__ == "__main__": | |
| main() | |