PanoGS / panogs /rendering /cpu /rasterizer.py
SiddhuG01's picture
Deploy PanoGS Gradio proof of concept
42bed50
Raw History Blame Contribute Delete
8.65 kB
"""
CPU Gaussian Splatting Rasterizer.
Implements perspective Jacobian projection, 2D covariance calculation,
depth sorting, and front-to-back alpha compositing in vectorized NumPy.
"""
from typing import Optional, Tuple
import numpy as np
from panogs.core.gaussian.model import GaussianModel
from panogs.core.logging import get_logger
from panogs.rendering.camera import Camera
def project_gaussians_to_2d(
model: GaussianModel,
camera: Camera,
min_depth: float = 0.05,
low_pass_filter: float = 0.3,
) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
"""
Project 3D Gaussians into 2D screen space:
1. World-to-Camera coordinate transformation
2. Frustum near-plane culling (tz > min_depth)
3. Pinhole screen projection
4. Jacobian affine approximation: Sigma_2D = J * W * Sigma_3D * W^T * J^T + low_pass * I
Args:
model: GaussianModel instance.
camera: Camera instance.
min_depth: Near plane cutoff in meters.
low_pass_filter: Filter added to 2D covariance diagonal for subpixel antialiasing.
Returns:
Tuple:
- visible_indices: (M,) indices of visible Gaussians.
- depths: (M,) camera-space Z depths (tz).
- screen_pts: (M, 2) 2D screen coordinates (xs, ys).
- cov2d: (M, 2, 2) 2D screen covariance matrices.
- radii: (M,) bounding box radii in pixels.
"""
xyz = model.get_xyz()
cov3d = model.get_covariance_3d()
# 1. Transform positions to camera space: t = R @ p + t_cam
p_cam = camera.world_to_camera(xyz) # (N, 3)
tx = p_cam[:, 0]
ty = p_cam[:, 1]
tz = p_cam[:, 2]
# 2. Near-plane culling
visible_mask = tz > min_depth
if not np.any(visible_mask):
return (
np.empty(0, dtype=np.int64),
np.empty(0, dtype=np.float32),
np.empty((0, 2), dtype=np.float32),
np.empty((0, 2, 2), dtype=np.float32),
np.empty(0, dtype=np.float32),
)
vis_idx = np.nonzero(visible_mask)[0]
tx = tx[vis_idx]
ty = ty[vis_idx]
tz = tz[vis_idx]
cov3d_vis = cov3d[vis_idx]
fx = camera.fx
fy = camera.fy
cx = camera.cx
cy = camera.cy
# 3. Perspective projection
xs = (fx * tx / tz) + cx
ys = (fy * ty / tz) + cy
screen_pts = np.stack([xs, ys], axis=-1).astype(np.float32)
# 4. Transform 3D covariance to camera view frame: Sigma_cam = R_view @ Sigma_3D @ R_view^T
# R_view: (3, 3)
R_view = camera.R # (3, 3)
# cov3d_cam = R_view @ cov3d_vis @ R_view^T
cov3d_cam = np.matmul(R_view[np.newaxis, :, :], np.matmul(cov3d_vis, R_view.T[np.newaxis, :, :]))
# 5. Jacobian of projection J: (M, 2, 3)
M = len(vis_idx)
tz2 = tz * tz
J = np.zeros((M, 2, 3), dtype=np.float32)
J[:, 0, 0] = fx / tz
J[:, 0, 2] = - (fx * tx) / tz2
J[:, 1, 1] = fy / tz
J[:, 1, 2] = - (fy * ty) / tz2
# 6. Compute 2D covariance: Sigma_2D = J @ cov3d_cam @ J^T
cov2d = np.matmul(J, np.matmul(cov3d_cam, J.transpose(0, 2, 1)))
cov2d = np.nan_to_num(cov2d, nan=0.0, posinf=1000.0, neginf=-1000.0)
# Add low-pass filter to diagonal
cov2d[:, 0, 0] += low_pass_filter
cov2d[:, 1, 1] += low_pass_filter
# 7. Compute bounding box radii (3 * sqrt(max(cov_xx, cov_yy)))
r_x = 3.0 * np.sqrt(np.clip(cov2d[:, 0, 0], 1e-4, 10000.0))
r_y = 3.0 * np.sqrt(np.clip(cov2d[:, 1, 1], 1e-4, 10000.0))
radii = np.clip(np.maximum(r_x, r_y), 0.1, max(camera.width, camera.height) * 2.0).astype(np.float32)
return vis_idx, tz, screen_pts, cov2d, radii
def render_gaussians_cpu(
model: GaussianModel,
camera: Camera,
bg_color: Tuple[float, float, float] = (1.0, 1.0, 1.0),
min_depth: float = 0.05,
) -> np.ndarray:
"""
Render 3D Gaussians from the viewpoint of `camera` on CPU using front-to-back alpha compositing.
Args:
model: GaussianModel containing 3D Gaussians.
camera: Viewpoint camera.
bg_color: RGB background color in [0, 1] (default: white (1, 1, 1)).
min_depth: Near plane clipping distance.
Returns:
np.ndarray: (H, W, 3) uint8 RGB rendered image.
"""
logger = get_logger("rendering.cpu")
H, W = camera.height, camera.width
# 1. Project Gaussians to 2D screen space
vis_idx, depths, screen_pts, cov2d, radii = project_gaussians_to_2d(
model, camera, min_depth=min_depth
)
if len(vis_idx) == 0:
logger.warning("No Gaussians visible in current camera frustum.")
bg_rgb = (np.array(bg_color, dtype=np.float32) * 255.0).astype(np.uint8)
return np.full((H, W, 3), bg_rgb, dtype=np.uint8)
# 2. Front-to-back depth sorting (ascending order of camera depth tz)
sort_order = np.argsort(depths)
vis_idx = vis_idx[sort_order]
screen_pts = screen_pts[sort_order]
cov2d = cov2d[sort_order]
radii = radii[sort_order]
# Pre-fetch colors and opacities for sorted visible Gaussians
all_colors = model.get_rgb() # (N, 3) float [0, 1]
all_opacities = model.get_opacity().flatten() # (N,) float [0, 1]
sorted_colors = all_colors[vis_idx]
sorted_opacities = all_opacities[vis_idx]
# Precompute 2D covariance inverse determinants: det = a*c - b^2
a = cov2d[:, 0, 0]
b = cov2d[:, 0, 1]
c = cov2d[:, 1, 1]
dets = a * c - b * b
valid_dets = np.isfinite(dets) & (dets > 1e-6) & np.isfinite(screen_pts[:, 0]) & np.isfinite(screen_pts[:, 1]) & np.isfinite(radii)
# Buffers for alpha compositing
# Accumulated color C (H, W, 3) float32 in [0, 1]
acc_color = np.zeros((H, W, 3), dtype=np.float32)
# Remaining transmittance T (H, W) float32 in [0, 1]
transmittance = np.ones((H, W), dtype=np.float32)
M = len(vis_idx)
rendered_count = 0
# 3. Rasterize and alpha composite each Gaussian
for i in range(M):
if not valid_dets[i]:
continue
cx, cy = screen_pts[i]
rad = radii[i]
det = dets[i]
# Bounding box in integer pixel coordinates
u_min = max(0, min(W - 1, int(np.floor(cx - rad))))
u_max = min(W - 1, max(0, int(np.ceil(cx + rad))))
v_min = max(0, min(H - 1, int(np.floor(cy - rad))))
v_max = min(H - 1, max(0, int(np.ceil(cy + rad))))
if u_min > u_max or v_min > v_max:
continue
# Check if bounding box region is already fully saturated (transmittance near 0)
T_box = transmittance[v_min:v_max + 1, u_min:u_max + 1]
if np.max(T_box) < 1e-4:
continue
# Generate 2D pixel coordinate grid for the bounding box
u_range = np.arange(u_min, u_max + 1, dtype=np.float32)
v_range = np.arange(v_min, v_max + 1, dtype=np.float32)
u_grid, v_grid = np.meshgrid(u_range, v_range, indexing="xy")
dx = u_grid - cx
dy = v_grid - cy
# Inverse 2D covariance quadratic form:
# Delta^T * Sigma^-1 * Delta = (c * dx^2 - 2 * b * dx * dy + a * dy^2) / det
inv_a = a[i]
inv_b = b[i]
inv_c = c[i]
power = -0.5 * (inv_c * dx * dx - 2.0 * inv_b * dx * dy + inv_a * dy * dy) / det
# Keep points with power <= 0
inside = power <= 0.0
if not np.any(inside):
continue
# Gaussian density G(p) in [0, 1]
G = np.zeros_like(power)
G[inside] = np.exp(power[inside])
# Alpha = opacity * G
alpha = sorted_opacities[i] * G
# Skip subpixel contributions (< 1/255)
alpha_mask = alpha >= (1.0 / 255.0)
if not np.any(alpha_mask):
continue
# Over-operator compositing weight: w = alpha * T
w = alpha * T_box
color_i = sorted_colors[i] # (3,)
# Accumulate color: C += w * color_i
acc_color[v_min:v_max + 1, u_min:u_max + 1] += w[:, :, np.newaxis] * color_i[np.newaxis, np.newaxis, :]
# Update transmittance: T *= (1 - alpha)
transmittance[v_min:v_max + 1, u_min:u_max + 1] *= (1.0 - alpha)
rendered_count += 1
# 4. Composite remaining background
bg = np.array(bg_color, dtype=np.float32).reshape(1, 1, 3)
final_rgb = acc_color + transmittance[:, :, np.newaxis] * bg
linear_rgb = np.clip(final_rgb, 0.0, 1.0)
srgb = np.where(
linear_rgb <= 0.0031308,
12.92 * linear_rgb,
1.055 * np.power(linear_rgb, 1.0 / 2.4) - 0.055,
)
final_uint8 = (srgb * 255.0).round().astype(np.uint8)
logger.debug(f"Rendered {rendered_count:,} / {M:,} visible Gaussians to {W}x{H} frame.")
return final_uint8