scripts/demo_extrapolation.py

#!/usr/bin/env python3
"""Render a lateral-shift sweep for one scene and save a comparison grid:
rows = lateral shift, cols = [MapGS render | GT deviated view]. Showcases the
out-of-trajectory simulation use case (§4.8).

    python scripts/demo_extrapolation.py --config configs/synthetic_smoke.yaml \\
        --ckpt runs/mapgs_synth/ckpt_final.pt --scene 0 --out demo.png
"""

import argparse

import numpy as np
import torch

from mapgs.config import load_config
from mapgs.data import SyntheticDataset
from mapgs.data.synthetic import render_deviated_gt
from mapgs.eval.evaluator import Evaluator
from mapgs.losses import perturb_pose
from mapgs.model import MapGS
from mapgs.model.dynamic import place_dynamic_gaussians


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--config", type=str, default="configs/synthetic_smoke.yaml")
    ap.add_argument("--override", nargs="*", default=[])
    ap.add_argument("--ckpt", type=str, default=None)
    ap.add_argument("--scene", type=int, default=0)
    ap.add_argument("--frame", type=int, default=None)
    ap.add_argument("--out", type=str, default="demo_extrapolation.png")
    args = ap.parse_args()

    import imageio.v2 as imageio

    cfg = load_config(args.config, args.override)
    model = MapGS(cfg).to(cfg.device)
    if args.ckpt:
        model.load_state_dict(torch.load(args.ckpt, map_location=cfg.device, weights_only=False)["model"])
    model.eval()

    ds = SyntheticDataset(cfg, "train", device=cfg.device)
    ev = Evaluator(model, cfg, device=cfg.device)
    s = ds[args.scene]
    scene = ds.get_scene(args.scene)
    frame = args.frame if args.frame is not None else cfg.data.num_frames // 2
    g, dyn = ev._decode(s)
    g_f = g if dyn is None else place_dynamic_gaussians(
        g, dyn["box_centers"], dyn["box_rots"], dyn["canon_idx"], frame)
    Kc = scene.K[1].to(cfg.device)
    base = scene.cam2world[frame, 1].to(cfg.device)

    rows = []
    for sh in cfg.eval.lateral_shifts:
        dev = perturb_pose(base, lateral=float(sh))
        out = ev.ras.render(g_f, Kc[None], dev[None], cfg.data.height, cfg.data.width)
        pred = (model.feature_to_rgb(out.color) if model.uses_features else out.color[:, :3].clamp(0, 1))[0]
        gt, _ = render_deviated_gt(scene, frame, dev.cpu(), cfg.device, cfg.data.height, cfg.data.width, ev.ras)
        row = torch.cat([pred.cpu(), gt], dim=-1)  # [3, H, 2W]
        rows.append(row)
    grid = torch.cat(rows, dim=-2).permute(1, 2, 0).clamp(0, 1).numpy()
    imageio.imwrite(args.out, (grid * 255).astype(np.uint8))
    print(f"wrote {args.out} (rows = shifts {list(cfg.eval.lateral_shifts)} m; left=MapGS, right=GT)")


if __name__ == "__main__":
    main()