"""Q3.3: compare direct point-head output with points unprojected from predicted camera + depth, against GT canonical points (normalized units). Run from the repository root: python submission/compare_pointmaps.py --split val --out submission/pointmap_val.json """ import argparse import json import torch from a4data.dataset import vggt_targets from a4data.loading import make_dataloader from models.vggt.runtime import add_arguments, build_components, load_config, predict CHECKPOINTS = { "baseline (camera+depth)": "outputs/vggt_train2/best.pt", "point model": "outputs/vggt_points_b4w2_2/best.pt", } def errors_for(pred_points, gt_points, valid): """pred/gt: (V,3,H,W), valid: (V,1,H,W) -> per-pixel L2 errors and GT norms.""" m = valid[:, 0] & torch.isfinite(pred_points).all(1) err = (pred_points - gt_points).norm(dim=1)[m] return err, gt_points.norm(dim=1)[m] def summarize(errs, norms): e, n = torch.cat(errs), torch.cat(norms) return { "mean_l2": e.mean().item(), "median_l2": e.median().item() if e.numel() < 2**24 else e[::4].median().item(), "rel_l2": (e / n.clamp_min(1e-6)).mean().item(), "acc@0.05": (e < 0.05).float().mean().item(), "acc@0.10": (e < 0.10).float().mean().item(), "num_pixels": e.numel(), } @torch.no_grad() def main(): parser = argparse.ArgumentParser(description=__doc__) add_arguments(parser) parser.add_argument("--split", choices=["val", "test"], default="val") parser.add_argument("--out") args = parser.parse_args() report = {"split": args.split} for name, path in CHECKPOINTS.items(): ckpt = torch.load(path, map_location="cpu", weights_only=True) config = load_config(args, ckpt["config"]) model, _, geometry = build_components(config, from_checkpoint=True) model.load_state_dict(ckpt["model"]) model.to(args.device).eval() loader = make_dataloader(config["data"], args.split, geometry) acc = {"camera+depth": ([], []), "point head": ([], [])} for batch in loader: pred = {k: v.float().cpu() if v.is_floating_point() else v.cpu() for k, v in predict(model, batch, args.device).items()} tgt = vggt_targets(batch, include_points=True) for b in range(batch["rgb"].shape[0]): gt, valid = tgt["points_world"][b], tgt["depth_valid"][b].bool() unproj = geometry.unproject_depth( pred["K"][b], pred["T_world_from_camera"][b], pred["depth"][b] ).permute(0, 3, 1, 2) e, n = errors_for(unproj, gt, valid) acc["camera+depth"][0].append(e); acc["camera+depth"][1].append(n) if "points_world" in pred: e, n = errors_for(pred["points_world"][b], gt, valid) acc["point head"][0].append(e); acc["point head"][1].append(n) for path_name, (errs, norms) in acc.items(): if errs: report[f"{name} / {path_name}"] = summarize(errs, norms) del model torch.cuda.empty_cache() text = json.dumps(report, indent=2) print(text) if args.out: open(args.out, "w").write(text + "\n") if __name__ == "__main__": main()