import os import cv2 import torch import numpy as np # ---- paths (edit these) ---- OUT_DIR = "./outputs/demo/test_10" VIDEO_PATH = os.path.join(OUT_DIR, "1_incam.mp4") BBX_PATH = os.path.join(OUT_DIR, "preprocess", "bbx.pt") VITPOSE_PATH = os.path.join(OUT_DIR, "preprocess", "vitpose.pt") OUT_BBOX_ONLY = os.path.join(OUT_DIR, "debug_bbox_only_on_incam.mp4") OUT_BBOX_KP = os.path.join(OUT_DIR, "debug_bbox_kp_on_incam.mp4") # --------------------------- COCO17_NAMES = [ "nose", "l_eye", "r_eye", "l_ear", "r_ear", "l_sho", "r_sho", "l_elb", "r_elb", "l_wri", "r_wri", "l_hip", "r_hip", "l_knee", "r_knee", "l_ank", "r_ank" ] def to_numpy(x): if isinstance(x, torch.Tensor): return x.detach().cpu().numpy() return np.array(x) def xyxy_to_xys(bbx_xyxy_t: torch.Tensor) -> torch.Tensor: """(L,4) xyxy -> (L,3) (cx,cy,s) where s is square side = max(w,h).""" x1, y1, x2, y2 = bbx_xyxy_t.unbind(-1) cx = (x1 + x2) * 0.5 cy = (y1 + y2) * 0.5 w = (x2 - x1).clamp(min=1.0) h = (y2 - y1).clamp(min=1.0) s = torch.maximum(w, h) return torch.stack([cx, cy, s], dim=-1) def xys_to_xyxy(bbx_xys_t: torch.Tensor) -> torch.Tensor: """(L,3) (cx,cy,s) -> (L,4) xyxy of square.""" cx, cy, s = bbx_xys_t.unbind(-1) hs = s * 0.5 x1 = cx - hs y1 = cy - hs x2 = cx + hs y2 = cy + hs return torch.stack([x1, y1, x2, y2], dim=-1) def draw_bbox_xyxy(frame, xyxy, color=(0, 255, 0), thickness=2): x1, y1, x2, y2 = [int(round(float(v))) for v in xyxy] cv2.rectangle(frame, (x1, y1), (x2, y2), color, thickness) return frame def draw_kps_with_names(frame, kps_xy, conf=None, radius=3, show_conf=True, conf_thr=0.0): """ kps_xy: (J,2) in image pixels conf: (J,) optional - Always renders the label (including confidence) so you can verify confidence behavior. - Text is BLACK with a white background box for readability. """ H, W = frame.shape[:2] for j, (x, y) in enumerate(kps_xy): x_i, y_i = int(round(float(x))), int(round(float(y))) if x_i < 0 or y_i < 0 or x_i >= W or y_i >= H: continue # Confidence if conf is None: c = 1.0 else: c = float(conf[j]) ok = (c >= conf_thr) # Draw joint marker (keep your preferred colors; this is just for visibility) if conf is None: pt_color = (0, 0, 255) # red else: pt_color = (0, 0, 255) if ok else (150, 150, 150) # red if ok, gray if low cv2.circle(frame, (x_i, y_i), radius, pt_color, -1) # Label name = COCO17_NAMES[j] if j < len(COCO17_NAMES) else f"j{j}" label = f"{name} {c:.2f}" if (conf is not None and show_conf) else name # Position label org = (x_i + 4, y_i - 6) # Compute text size for background box font = cv2.FONT_HERSHEY_SIMPLEX font_scale = 0.40 thickness = 1 (tw, th), baseline = cv2.getTextSize(label, font, font_scale, thickness) # Background rectangle (white), then black text on top x0, y0 = org[0], org[1] - th x1, y1 = org[0] + tw, org[1] + baseline # Clamp background box within frame x0 = max(0, min(W - 1, x0)) y0 = max(0, min(H - 1, y0)) x1 = max(0, min(W - 1, x1)) y1 = max(0, min(H - 1, y1)) cv2.rectangle(frame, (x0, y0), (x1, y1), (255, 255, 255), -1) # filled white cv2.putText(frame, label, org, font, font_scale, (0, 0, 0), thickness, cv2.LINE_AA) # black text return frame def convert_kp_to_image_pixels(kp, bbx_xys, crop_size=256): """ Convert kp to full-image pixel coords using HMR-style crop mapping: x_img = cx + x_norm * (s/2) y_img = cy + y_norm * (s/2) Supports: - kp in [-1,1] (normalized crop coords) - kp in crop pixels [0..crop_size-1] - kp already in image pixels (then returned unchanged) """ kp = np.asarray(kp, dtype=np.float32) # (L,J,2) bbx_xys = np.asarray(bbx_xys, dtype=np.float32) # (L,3) kp_min = float(np.nanmin(kp)) kp_max = float(np.nanmax(kp)) # Decide mode if kp_min >= -1.5 and kp_max <= 1.5: mode = "norm_pm1" # [-1,1] kp_norm = kp elif kp_min >= -5.0 and kp_max <= (crop_size + 5.0): mode = "crop_pixels" # crop pixels -> [-1,1] denom = (crop_size - 1.0) kp_norm = (kp / denom) * 2.0 - 1.0 else: mode = "image_pixels" return kp, mode, (kp_min, kp_max) cx = bbx_xys[:, 0:1] # (L,1) cy = bbx_xys[:, 1:2] # (L,1) s = bbx_xys[:, 2:3] # (L,1) hs = s * 0.5 x_img = cx + kp_norm[..., 0] * hs y_img = cy + kp_norm[..., 1] * hs kp_img = np.stack([x_img, y_img], axis=-1) return kp_img, mode, (kp_min, kp_max) def main(): # ---- Load bbox ---- bbx = torch.load(BBX_PATH, map_location="cpu") bbx_xyxy_t = bbx.get("bbx_xyxy", None) bbx_xys_t = bbx.get("bbx_xys", None) if bbx_xyxy_t is None and bbx_xys_t is None: raise ValueError("bbx.pt must contain 'bbx_xyxy' and/or 'bbx_xys'.") if bbx_xys_t is None and bbx_xyxy_t is not None: bbx_xys_t = xyxy_to_xys(bbx_xyxy_t) if bbx_xyxy_t is None and bbx_xys_t is not None: bbx_xyxy_t = xys_to_xyxy(bbx_xys_t) bbx_xyxy = to_numpy(bbx_xyxy_t) # (L,4) bbx_xys = to_numpy(bbx_xys_t) # (L,3) print("bbx_xyxy shape:", bbx_xyxy.shape, "bbx_xys shape:", bbx_xys.shape) # ---- Load vitpose ---- vitpose = torch.load(VITPOSE_PATH, map_location="cpu") conf = None if isinstance(vitpose, dict): kp = None for k in ["kp2d", "keypoints", "kps", "joints_2d", "vitpose"]: if k in vitpose: kp = vitpose[k] break if kp is None: print("vitpose.pt keys:", list(vitpose.keys())) raise ValueError("Couldn't find keypoints in vitpose dict.") kp = to_numpy(kp) for k in ["conf", "confidence", "scores", "kp2d_conf", "keypoint_scores"]: if k in vitpose: conf = to_numpy(vitpose[k]) break else: kp = to_numpy(vitpose) if kp.ndim != 3: raise ValueError(f"Unexpected kp shape: {kp.shape} (expected L x J x 2/3)") if kp.shape[-1] == 3 and conf is None: conf = kp[..., 2] kp = kp[..., :2] elif kp.shape[-1] != 2: raise ValueError(f"Unexpected kp last dim: {kp.shape[-1]} (expected 2 or 3)") # ---- Open video ---- cap = cv2.VideoCapture(VIDEO_PATH) if not cap.isOpened(): raise RuntimeError(f"Failed to open video: {VIDEO_PATH}") fps = cap.get(cv2.CAP_PROP_FPS) or 30.0 W = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)) H = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) print("Video:", VIDEO_PATH, "W,H:", W, H, "fps:", fps) # ---- Align lengths ---- L = min(len(bbx_xyxy), len(bbx_xys), kp.shape[0]) print("Using L =", L) # ---- Convert keypoints to image pixels if needed ---- kp_img, mode, (kp_min, kp_max) = convert_kp_to_image_pixels(kp[:L], bbx_xys[:L], crop_size=256) print(f"kp stats min/max: {kp_min:.3f} / {kp_max:.3f} -> interpreted as mode: {mode}") # Basic bbox sanity centers = np.stack( [(bbx_xyxy[:L, 0] + bbx_xyxy[:L, 2]) * 0.5, (bbx_xyxy[:L, 1] + bbx_xyxy[:L, 3]) * 0.5], axis=-1 ) center_speed = np.linalg.norm(centers[1:] - centers[:-1], axis=-1) if len(center_speed) > 0: print("bbox center jump px (p50/p90/max):", float(np.percentile(center_speed, 50)), float(np.percentile(center_speed, 90)), float(center_speed.max())) if conf is not None: conf_use = conf[:L] print("kp conf (mean/p10):", float(np.mean(conf_use)), float(np.percentile(conf_use, 10))) # ---- Writers ---- fourcc = cv2.VideoWriter_fourcc(*"mp4v") w_bbox = cv2.VideoWriter(OUT_BBOX_ONLY, fourcc, fps, (W, H)) w_kp = cv2.VideoWriter(OUT_BBOX_KP, fourcc, fps, (W, H)) t = 0 while t < L: ok, frame = cap.read() if not ok: break f1 = frame.copy() f2 = frame.copy() draw_bbox_xyxy(f1, bbx_xyxy[t]) draw_bbox_xyxy(f2, bbx_xyxy[t]) c_t = conf[t] if conf is not None else None draw_kps_with_names(f2, kp_img[t], conf=c_t, show_conf=True) cv2.putText(f1, f"t={t}", (10, 25), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (255, 255, 255), 2, cv2.LINE_AA) cv2.putText(f2, f"t={t} mode={mode}", (10, 25), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (255, 255, 255), 2, cv2.LINE_AA) w_bbox.write(f1) w_kp.write(f2) t += 1 cap.release() w_bbox.release() w_kp.release() print("Saved:", OUT_BBOX_ONLY) print("Saved:", OUT_BBOX_KP) if __name__ == "__main__": main()