code/realscene.py

"""E12: real-world scene (pretrained 3DGS) generalization + renders.

Loads a pretrained 3D Gaussian Splatting checkpoint (Inria .ply format), renders
it from synthesized orbit cameras, and repeats the criticality and support-guard
measurement on it.  This checks that the concentration of risk and the guard
generalize beyond the trained synthetic scenes, and leaves high-resolution renders
(clean / faulted / guarded) for the manuscript.
"""
import argparse
import json
import os

import numpy as np
import torch
from plyfile import PlyData
import imageio.v2 as imageio

import faultlib as F
import gsmodel


def load_ply(path):
    ply = PlyData.read(path)["vertex"]
    xyz = np.stack([ply["x"], ply["y"], ply["z"]], 1).astype(np.float32)
    opac = np.asarray(ply["opacity"], np.float32)[:, None]
    scales = np.stack([ply[f"scale_{i}"] for i in range(3)], 1).astype(np.float32)
    rots = np.stack([ply[f"rot_{i}"] for i in range(4)], 1).astype(np.float32)
    fdc = np.stack([ply[f"f_dc_{i}"] for i in range(3)], 1).astype(np.float32)  # [N,3]
    rest_names = sorted([p.name for p in ply.properties if p.name.startswith("f_rest_")],
                        key=lambda s: int(s.split("_")[-1]))
    frest = np.stack([ply[n] for n in rest_names], 1).astype(np.float32)  # [N, 3*M]
    N = xyz.shape[0]
    M = frest.shape[1] // 3
    sh_rest = frest.reshape(N, 3, M).transpose(0, 2, 1)  # [N, M, 3]
    sh_deg = int(round((M + 1) ** 0.5)) - 1
    dev = "cuda"
    params = {
        "means": torch.tensor(xyz, device=dev),
        "scales": torch.tensor(scales, device=dev),
        "quats": torch.tensor(rots, device=dev),
        "opacities": torch.tensor(opac.squeeze(1), device=dev),
        "sh0": torch.tensor(fdc, device=dev).reshape(N, 1, 3),
        "shN": torch.tensor(sh_rest, device=dev),
    }
    return params, sh_deg, N


def orbit_cameras(means, n_views, W, H, fov_deg=60.0):
    # robust center/extent: ignore distant floaters via the median and a low quantile
    c = means.median(0).values
    X = (means - c)
    d = X.norm(dim=1)
    keep = d < torch.quantile(d, 0.6)        # dense core only
    Xc = X[keep]
    cov = (Xc.T @ Xc) / Xc.shape[0]
    evals, evecs = torch.linalg.eigh(cov)
    up = evecs[:, 0]
    a1 = evecs[:, 2]; a2 = evecs[:, 1]       # ground-plane axes
    radius = float(torch.quantile(d, 0.5).item()) * 1.2
    elev = radius * 0.12
    vms = []
    for i in range(n_views):
        th = 2 * np.pi * i / n_views
        cam = c + radius * (np.cos(th) * a1 + np.sin(th) * a2) + elev * up
        z = (c - cam); z = z / z.norm()              # forward (+z, OpenCV)
        x = torch.linalg.cross(z, up); x = x / x.norm()
        y = torch.linalg.cross(z, x)                  # down
        R = torch.stack([x, y, z], 0)                 # world->cam rotation (rows)
        t = -R @ cam
        vm = torch.eye(4, device=means.device)
        vm[:3, :3] = R; vm[:3, 3] = t
        vms.append(vm)
    vms = torch.stack(vms, 0)
    f = 0.5 * W / np.tan(0.5 * np.deg2rad(fov_deg))
    K = torch.tensor([[f, 0, W / 2], [0, f, H / 2], [0, 0, 1]], device=means.device, dtype=torch.float32)
    Ks = K[None].repeat(n_views, 1, 1)
    return vms, Ks


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--ply", required=True)
    ap.add_argument("--name", default="truck")
    ap.add_argument("--out", default="/root/seu/results/realscene")
    ap.add_argument("--gen", default="/root/seu/results/generated")
    ap.add_argument("--W", type=int, default=800)
    ap.add_argument("--S", type=int, default=300)
    ap.add_argument("--views", type=int, default=8)
    args = ap.parse_args()
    os.makedirs(args.out, exist_ok=True); os.makedirs(args.gen, exist_ok=True)
    dev = "cuda"
    params, sh, N = load_ply(args.ply)
    W = H = args.W
    print(f"[{args.name}] N={N} sh_deg={sh}", flush=True)
    vms, Ks = orbit_cameras(params["means"], args.views, W, H)
    bounds = F.compute_bounds(params)
    stored, work = F.quantize_params(params, "fp32")

    # pretty clean renders
    clean, _ = F.render_views(work, vms, Ks, W, H, sh)
    for i in range(min(3, args.views)):
        imageio.imwrite(os.path.join(args.out, f"{args.name}_view{i}.png"),
                        (clean[i].cpu().numpy() * 255).astype(np.uint8))

    # a clean/faulted/guarded triptych on a chosen view, using a large-footprint scale-sign flip
    v = 0
    rng = np.random.default_rng(0)
    best = (-1, None)
    sc = work["scales"]
    for _ in range(150):
        g = int(rng.integers(0, N)); flat = g * 3
        cv, _ = F.flip_one(stored["scales"], sc, flat, 31, "fp32")
        img, _ = F.render_views(work, vms[v:v + 1], Ks[v:v + 1], W, H, sh)
        F.restore_one(sc, flat, cv)
        fp = ((img[0] - clean[v]).abs().amax(-1) > 1 / 255).float().mean().item()
        if fp > best[0]:
            best = (fp, g)
    g = best[1]; flat = g * 3
    cv, _ = F.flip_one(stored["scales"], sc, flat, 31, "fp32")
    faulted, _ = F.render_views(work, vms[v:v + 1], Ks[v:v + 1], W, H, sh)
    gw = F.apply_guard(work, bounds)
    guarded, _ = F.render_views(gw, vms[v:v + 1], Ks[v:v + 1], W, H, sh)
    F.restore_one(sc, flat, cv)
    trip = torch.cat([clean[v], faulted[0], guarded[0]], dim=1).clamp(0, 1)
    imageio.imwrite(os.path.join(args.gen, f"fig_realscene.png"),
                    (trip.cpu().numpy() * 255).astype(np.uint8))

    # criticality on the real scene, measured against a multi-view clean reference
    clean_small, _ = F.render_views(work, vms[:4], Ks[:4], W, H, sh)

    def footprint(rp):
        img, cat = F.render_views(rp, vms[:4], Ks[:4], W, H, sh)
        return ((img - clean_small).abs().amax(-1) > 1 / 255).float().mean().item(), bool(cat)

    # (a) dominant scale sign-bit case, no guard vs guard
    sign_ng, sign_g = [], []
    for _ in range(150):
        g = int(rng.integers(0, N)); flat = g * 3
        cv, _ = F.flip_one(stored["scales"], work["scales"], flat, 31, "fp32")
        f_ng, _ = footprint(work)
        f_g, _ = footprint(F.apply_guard(work, bounds))
        F.restore_one(work["scales"], flat, cv)
        sign_ng.append(f_ng); sign_g.append(f_g)
    # (b) overall uniform-random single-bit catastrophe rate, no guard vs guard
    FIELDS = ["means", "scales", "quats", "opacities", "sh0", "shN"]
    comps = {f: work[f].reshape(N, -1).shape[1] for f in FIELDS}
    cats_ng, cats_g = [], []
    for _ in range(args.S):
        field = FIELDS[int(rng.integers(0, 6))]
        flat = int(rng.integers(0, N * comps[field])); bit = int(rng.integers(0, 32))
        cv, _ = F.flip_one(stored[field], work[field], flat, bit, "fp32")
        f_ng, c_ng = footprint(work)
        f_g, c_g = footprint(F.apply_guard(work, bounds))
        F.restore_one(work[field], flat, cv)
        cats_ng.append(int(c_ng or f_ng > 0.01)); cats_g.append(int(c_g or f_g > 0.01))
    res = {"name": args.name, "N": int(N), "sh_degree": sh,
           "cat_rate_noguard": float(np.mean(cats_ng)), "cat_rate_guard": float(np.mean(cats_g)),
           "scalesign_foot_noguard": float(np.mean(sign_ng) * 100),
           "scalesign_foot_guard": float(np.mean(sign_g) * 100),
           "scalesign_p99_noguard": float(np.percentile(sign_ng, 99) * 100),
           "n_samples": args.S}
    json.dump(res, open(os.path.join(args.out, f"realscene_{args.name}.json"), "w"), indent=2)
    print("REALSCENE", res, flush=True)


if __name__ == "__main__":
    main()