| |
| """角色清晰度诊断·分析端：按 ids 里的角色掩码裁框，比较 GT(源 1280×704) / GT(832×480) / VAE 重建 / 各生成变体。 |
| 指标（框内，相对 GT832）：sharp = Sobel 梯度均值之比，lap = Laplacian 方差之比，psnr。另出无损裁剪对比图。""" |
| import json |
| import os |
| import sys |
|
|
| import cv2 |
| import numpy as np |
|
|
| import geometry as GEO |
| import model as GM |
|
|
| M = "/opt/dlami/nvme/zhiyangdeng/maliang" |
| D = f"{M}/outputs/char_diag" |
| VARIANTS = ["s4000_cfg1", "s8000_cfg1", "s8000_semcfg2", "s8000_semcfg3"] |
| ROLES = {"player": GM.PLAYER_PART_IDS, "boss": GM.BOSS_PART_IDS} |
| PAD, MIN_PX, UP = 10, 600, 3 |
|
|
|
|
| def bbox(mask): |
| ys, xs = np.nonzero(mask) |
| return max(0, ys.min() - PAD), min(mask.shape[0], ys.max() + PAD + 1), max(0, xs.min() - PAD), min(mask.shape[1], xs.max() + PAD + 1) |
|
|
|
|
| def grad(img): |
| g = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY).astype(np.float32) |
| gx, gy = cv2.Sobel(g, cv2.CV_32F, 1, 0), cv2.Sobel(g, cv2.CV_32F, 0, 1) |
| return float(np.sqrt(gx ** 2 + gy ** 2).mean()), float(cv2.Laplacian(g, cv2.CV_32F).var()) |
|
|
|
|
| def psnr(a, b): |
| mse = np.mean((a.astype(np.float32) - b.astype(np.float32)) ** 2) |
| return 10 * np.log10(255.0 ** 2 / max(mse, 1e-6)) |
|
|
|
|
| def src_crop(src, box): |
| """GT 源分辨率里与 832×480 框对应的区域，缩放到框的尺寸 ×UP（看 1280×704 原生能有多清楚）。""" |
| y0, y1, x0, x1 = box |
| p = GEO.dst_to_src(np.array([[x0 - 0.5, y0 - 0.5], [x1 - 0.5, y1 - 0.5]])) |
| sx0, sy0 = int(max(0, np.floor(p[0, 0]))), int(max(0, np.floor(p[0, 1]))) |
| sx1, sy1 = int(min(src.shape[1], np.ceil(p[1, 0]))), int(min(src.shape[0], np.ceil(p[1, 1]))) |
| return cv2.resize(src[sy0:sy1, sx0:sx1], ((x1 - x0) * UP, (y1 - y0) * UP), interpolation=cv2.INTER_CUBIC) |
|
|
|
|
| def label(img, text): |
| out = img.copy() |
| cv2.rectangle(out, (0, 0), (out.shape[1], 18), (0, 0, 0), -1) |
| cv2.putText(out, text, (3, 13), cv2.FONT_HERSHEY_SIMPLEX, 0.42, (255, 230, 120), 1, cv2.LINE_AA) |
| return out |
|
|
|
|
| def main(): |
| clips = json.load(open(f"{D}/clips.json")) |
| variants = [v for v in VARIANTS if all(os.path.exists(f"{D}/{v}/{c}.npz") for c in clips)] |
| stats = {r: {k: [] for k in ["recon"] + variants} for r in ROLES} |
| sheet_rows = [] |
| for ci, cid in enumerate(clips): |
| ref = np.load(f"{D}/_ref/{cid}.npz") |
| gt, rec, src, ids = ref["gt"], ref["recon"], ref["gt_src"], ref["ids"] |
| gens = {v: np.load(f"{D}/{v}/{cid}.npz")["frames"] for v in variants} |
| for role, pids in ROLES.items(): |
| for t in range(8, 81, 8): |
| m = np.isin(ids[t], pids) |
| if m.sum() < MIN_PX: |
| continue |
| y0, y1, x0, x1 = box = bbox(m) |
| g_ref, l_ref = grad(gt[t, y0:y1, x0:x1]) |
| for k, fr in [("recon", rec)] + list(gens.items()): |
| c = fr[t, y0:y1, x0:x1] |
| g, l = grad(c) |
| stats[role][k].append((g / max(g_ref, 1e-6), l / max(l_ref, 1e-6), psnr(c, gt[t, y0:y1, x0:x1]))) |
| if role == "player" and t in (40, 72) and len(sheet_rows) < 12: |
| up = lambda a: cv2.resize(a, ((x1 - x0) * UP, (y1 - y0) * UP), interpolation=cv2.INTER_NEAREST) |
| tiles = [label(src_crop(src[t], box), "GT 1280x704 (native)"), label(up(gt[t, y0:y1, x0:x1]), "GT 832x480"), |
| label(up(rec[t, y0:y1, x0:x1]), "VAE recon")] |
| tiles += [label(up(gens[v][t, y0:y1, x0:x1]), v) for v in variants] |
| h = max(x.shape[0] for x in tiles) |
| tiles = [cv2.copyMakeBorder(x, 0, h - x.shape[0], 0, 4, cv2.BORDER_CONSTANT) for x in tiles] |
| sheet_rows.append(np.concatenate(tiles, 1)) |
| summary = {} |
| for role, d in stats.items(): |
| summary[role] = {k: {"n": len(v), "sharp_ratio": round(float(np.mean([x[0] for x in v])), 3), |
| "lapvar_ratio": round(float(np.mean([x[1] for x in v])), 3), |
| "psnr": round(float(np.mean([x[2] for x in v])), 2)} for k, v in d.items() if v} |
| json.dump(summary, open(f"{D}/summary.json", "w"), indent=1) |
| print(json.dumps(summary, indent=1)) |
| if sheet_rows: |
| w = max(r.shape[1] for r in sheet_rows) |
| rows = [cv2.copyMakeBorder(r, 0, 6, 0, w - r.shape[1], cv2.BORDER_CONSTANT) for r in sheet_rows] |
| sheet = np.concatenate(rows, 0) |
| cv2.imwrite(f"{D}/player_crops.png", sheet[:, :, ::-1]) |
| print("sheet", sheet.shape) |
|
|
|
|
| if __name__ == "__main__": |
| sys.exit(main()) |
|
|
