slothfulxtx
commented
3 weeks ago Hi, can you provide more details about which branch you got the segment error on? If you're using the pose branch, here's the render code I implement for testing.
import torch
import math
from diff_gauss_pose import GaussianRasterizationSettings, GaussianRasterizer
from utils.sh import eval_sh
def pose_render(viewpoint_camera, gaussians, cfg, bg_color, viewmatrix, scaling_modifier = 1.0, override_color = None, extra_attrs=None, debug=False):
"""
Render the scene.
Background tensor (bg_color) must be on GPU!
"""
# Create zero tensor. We will use it to make pytorch return gradients of the 2D (screen-space) means
screenspace_points = torch.zeros_like(gaussians.get_xyz, dtype=gaussians.get_xyz.dtype, requires_grad=True, device="cuda") + 0
try:
screenspace_points.retain_grad()
except:
pass
# Set up rasterization configuration
tanfovx = math.tan(viewpoint_camera.FoVx * 0.5)
tanfovy = math.tan(viewpoint_camera.FoVy * 0.5)
raster_settings = GaussianRasterizationSettings(
image_height=int(viewpoint_camera.image_height),
image_width=int(viewpoint_camera.image_width),
tanfovx=tanfovx,
tanfovy=tanfovy,
bg=bg_color,
scale_modifier=scaling_modifier,
projmatrix=viewpoint_camera.projection_matrix,
sh_degree=gaussians.active_sh_degree if hasattr(gaussians, 'active_sh_degree') else 0,
prefiltered=False,
enable_cov_grad=True,
enable_sh_grad=False,
debug=debug
)
rasterizer = GaussianRasterizer(raster_settings=raster_settings)
means3D = gaussians.get_xyz
means2D = screenspace_points
opacity = gaussians.get_opacity
# If precomputed 3d covariance is provided, use it. If not, then it will be computed from
# scaling / rotation by the rasterizer.
scales = None
rotations = None
cov3D_precomp = None
if cfg.compute_cov3D_python:
cov3D_precomp = gaussians.get_covariance(scaling_modifier)
else:
scales = gaussians.get_scaling
rotations = gaussians.get_rotation
# If precomputed colors are provided, use them. Otherwise, if it is desired to precompute colors
# from SHs in Python, do it. If not, then SH -> RGB conversion will be done by rasterizer.
shs = None
colors_precomp = None
if override_color is None:
if cfg.convert_SHs_python:
shs_view = gaussians.get_features.transpose(1, 2).view(-1, 3, (gaussians.max_sh_degree+1)**2)
dir_pp = (gaussians.get_xyz - viewpoint_camera.camera_center.repeat(gaussians.get_features.shape[0], 1))
dir_pp_normalized = dir_pp/dir_pp.norm(dim=1, keepdim=True)
sh2rgb = eval_sh(gaussians.active_sh_degree, shs_view, dir_pp_normalized)
colors_precomp = torch.clamp_min(sh2rgb + 0.5, 0.0)
else:
shs = gaussians.get_features
else:
colors_precomp = override_color
# Rasterize visible Gaussians to image, obtain their radii (on screen).
rendered_image, rendered_depth, rendered_norm, rendered_alpha, radii, extra = rasterizer(
means3D = means3D,
means2D = means2D,
shs = shs,
colors_precomp = colors_precomp,
opacities = opacity,
scales = scales,
rotations = rotations,
cov3Ds_precomp = cov3D_precomp,
extra_attrs=extra_attrs,
viewmatrix=viewmatrix)
# Those Gaussians that were frustum culled or had a radius of 0 were not visible.
# They will be excluded from value updates used in the splitting criteria.
return {"render": rendered_image,
"depth": rendered_depth,
"norm": rendered_norm,
"alpha": rendered_alpha,
"viewspace_points": screenspace_points,
"visibility_filter" : radii > 0,
"extra": extra,
"radii": radii}
Hi!
Is there anybody who also get segmentation fault error in rasterizer_impl.cu at line:
ImageState imgState = ImageState::fromChunk(img_chunkptr,width * height);I get the segmentation fault error before the function "fromChunk" starts to process, because even the first line inside the function does not executed.
The original rasterizer implementation works fine, but I need to optimize the image pose, so I need pose gradients.
Can someone provide me an example usage of the code? (A render.py code would be the best :) )
Thank you, in advance!