# ============================================================ # bev_renderer.py # ============================================================ # # ETHUKU CREATE PANNINOM? # ━━━━━━━━━━━━━━━━━━━━━━ # Numbers (heatmap tensors) yaarukum puriyaadhu. Recruiter oru PADAM # paakanum: "6 camera photo -> top-down map-la cars marked". # Ithu antha padathai varaiyum. # # MUNADI FILE ODA CONNECTION: # ━━━━━━━━━━━━━━━━━━━━━━━━━━ # Input : training/evaluate.py-oda decode_predictions output # + dataset._boxes_in_ego (ground truth) # Output: PNG / GIF -> README-la pogum # # INNER OPERATIONS: # ━━━━━━━━━━━━━━━━ # Matplotlib-la 2 panel: (idathu) 6 camera photos grid, # (valathu) top-down BEV - ego car center-la, boxes suthi. # Box varaiyurathu: center + w,l + yaw -> 4 corner points -> polygon. # # INPUT / OUTPUT: # ━━━━━━━━━━━━━━ # render_frame(images, gt_boxes, pred_boxes, out_path) # # EPADI USE AAGUM: # ━━━━━━━━━━━━━━━ # python -m visualization.bev_renderer --ckpt runs/simplebev/best.pth # # ============================================================ import argparse import os import matplotlib matplotlib.use("Agg") # display illa (headless) - file-la mattum save import matplotlib.pyplot as plt import numpy as np from matplotlib.patches import Polygon from data.scripts.constants import ( CAMERAS, CLASSES, X_RANGE, Y_RANGE, IMAGENET_MEAN, IMAGENET_STD, ) # Ovvoru class-kum oru color. Recruiter oru nodi-la puriyanum. CLASS_COLORS = { 0: "#00d4ff", # car - cyan 1: "#ff8c00", # truck - orange 2: "#ffd700", # bus - gold 3: "#8b4513", # trailer - brown 4: "#9370db", # construction - purple 5: "#ff1493", # pedestrian - pink 6: "#00ff7f", # motorcycle - spring green 7: "#7fff00", # bicycle - chartreuse 8: "#ff4500", # traffic_cone - red-orange 9: "#a9a9a9", # barrier - grey } def denormalize(img: np.ndarray) -> np.ndarray: """ Normalize panna image-ai thirumba paakkura maadhiri maathurathu. Args: img: [3, H, W] normalized (-2.5 to 2.5) Returns: [H, W, 3] uint8-range float (0..1) - matplotlib ku """ img = img.transpose(1, 2, 0) # CHW -> HWC img = img * np.array(IMAGENET_STD) + np.array(IMAGENET_MEAN) return np.clip(img, 0, 1) def box_corners(x: float, y: float, w: float, l: float, yaw: float) -> np.ndarray: """ Box center + size + angle -> 4 mooli (corner) points. Math simple-a: mudhalla car origin-la irukku nu nenachi 4 corner kanakku podurom, appuram yaw angle-la suthi, appuram (x,y) ku nagarthurom. Args: x, y : center metres (x=munnadi, y=idathu) w, l : width, length metres yaw : radians Returns: [4, 2] corner coordinates """ # Car-oda soontha frame-la 4 corners (length = munnadi axis) corners = np.array([ [ l / 2, w / 2], [ l / 2, -w / 2], [-l / 2, -w / 2], [-l / 2, w / 2], ]) # 2D rotation matrix - yaw angle-la suthurathu c, s = np.cos(yaw), np.sin(yaw) R = np.array([[c, -s], [s, c]]) return corners @ R.T + np.array([x, y]) def draw_bev(ax, gt_boxes: list, pred_boxes: list, score_thresh: float = 0.2) -> None: """ Top-down BEV panel varaiyurathu. GT = white dashed outline (nijam) Pred = class color filled (namma model sonnathu) """ ax.set_facecolor("#0a0a14") # --- Grid lines (10m ku oru vari) - dooram purinjukka --- for r in range(10, 51, 10): ax.add_patch(plt.Circle((0, 0), r, fill=False, edgecolor="#2a2a3a", linewidth=0.6, zorder=1)) ax.text(r * 0.7, r * 0.7, f"{r}m", color="#4a4a5a", fontsize=7, zorder=1) # --- Ego car (namma vandi) center-la --- ego = box_corners(0, 0, 1.8, 4.5, 0) ax.add_patch(Polygon(ego, closed=True, facecolor="#ffffff", edgecolor="#ffffff", linewidth=1, zorder=5)) ax.arrow(0, 2.5, 0, 3, head_width=1.2, head_length=1.2, fc="#ffffff", ec="#ffffff", zorder=5) # munnadi direction # --- Ground truth --- for b in gt_boxes: pts = box_corners(b["x"], b["y"], b["w"], b["l"], b["yaw"]) ax.add_patch(Polygon(pts, closed=True, fill=False, edgecolor="#ffffff", linewidth=1.0, linestyle="--", alpha=0.55, zorder=3)) # --- Predictions --- n_drawn = 0 for b in pred_boxes: if b["score"] < score_thresh: continue color = CLASS_COLORS.get(b["cls"], "#ffffff") pts = box_corners(b["x"], b["y"], b["w"], b["l"], b["yaw"]) ax.add_patch(Polygon(pts, closed=True, facecolor=color, alpha=0.45, edgecolor=color, linewidth=1.2, zorder=4)) n_drawn += 1 ax.set_xlim(Y_RANGE[1], Y_RANGE[0]) # y = idathu, plot-la idathu pakkam ax.set_ylim(X_RANGE[0], X_RANGE[1]) # x = munnadi, plot-la mela ax.set_aspect("equal") ax.set_xlabel("left <- y (m) -> right", color="#8a8a9a", fontsize=8) ax.set_ylabel("back <- x (m) -> front", color="#8a8a9a", fontsize=8) ax.set_title(f"BEV | white dashed = ground truth | filled = predicted " f"({n_drawn} boxes, score>{score_thresh})", color="#e0e0e0", fontsize=9) ax.tick_params(colors="#4a4a5a", labelsize=7) for spine in ax.spines.values(): spine.set_color("#2a2a3a") def render_frame(images: np.ndarray, gt_boxes: list, pred_boxes: list, out_path: str, score_thresh: float = 0.2) -> None: """ Oru frame-oda full visualization save pannurathu. Args: images: [6, 3, 224, 400] normalized gt_boxes / pred_boxes: dict list (x, y, w, l, yaw, cls, [score]) out_path: PNG path """ fig = plt.figure(figsize=(17, 7), facecolor="#0a0a14") gs = fig.add_gridspec(2, 5, width_ratios=[1, 1, 1, 0.12, 1.9], hspace=0.08, wspace=0.04) # --- Idathu pakkam: 6 camera photos --- # Car-la epdi irukko appdiye layout: mela munnadi, keezha pinnadi. # Ippdi potta padathai paakkurava-ku surround view neradiya purinjidum. layout = [ ("CAM_FRONT_LEFT", 0, 0), ("CAM_FRONT", 0, 1), ("CAM_FRONT_RIGHT", 0, 2), ("CAM_BACK_LEFT", 1, 0), ("CAM_BACK", 1, 1), ("CAM_BACK_RIGHT", 1, 2), ] for cam, r, c in layout: ax = fig.add_subplot(gs[r, c]) ax.imshow(denormalize(images[CAMERAS.index(cam)])) ax.set_title(cam.replace("CAM_", ""), color="#8a8a9a", fontsize=7, pad=2) ax.axis("off") # --- Valathu pakkam: BEV --- ax_bev = fig.add_subplot(gs[:, 4]) draw_bev(ax_bev, gt_boxes, pred_boxes, score_thresh) # --- Class color legend --- present = sorted({b["cls"] for b in pred_boxes if b["score"] >= score_thresh} | {b["cls"] for b in gt_boxes}) handles = [plt.Line2D([0], [0], marker="s", linestyle="", markersize=7, color=CLASS_COLORS[c], label=CLASSES[c]) for c in present] if handles: ax_bev.legend(handles=handles, loc="upper left", fontsize=7, facecolor="#14141e", edgecolor="#2a2a3a", labelcolor="#c0c0d0") os.makedirs(os.path.dirname(out_path) or ".", exist_ok=True) fig.savefig(out_path, dpi=130, bbox_inches="tight", facecolor="#0a0a14") plt.close(fig) print(f"saved {out_path}") def main() -> None: ap = argparse.ArgumentParser() ap.add_argument("--ckpt", default="runs/simplebev/best.pth") ap.add_argument("--data-root", default="data/nuscenes-mini") ap.add_argument("--out-dir", default="runs/viz") ap.add_argument("--frames", type=int, default=6) ap.add_argument("--score-thresh", type=float, default=0.2) ap.add_argument("--gif", action="store_true", help="frames-ai GIF-a sethu vai") args = ap.parse_args() import torch from data.scripts.dataset import NuScenesBEVDataset from models.simplebev import SimpleBEV from training.evaluate import decode_predictions device = torch.device("mps" if torch.backends.mps.is_available() else "cpu") model = SimpleBEV(pretrained=False).to(device).eval() model.load_state_dict(torch.load(args.ckpt, map_location=device)["model"]) ds = NuScenesBEVDataset(args.data_root, split="val") paths = [] for i in range(min(args.frames, len(ds))): s = ds[i] with torch.no_grad(): preds = model(s["images"].unsqueeze(0).to(device), s["intrinsics"].unsqueeze(0).to(device), s["extrinsics"].unsqueeze(0).to(device)) pred_boxes = decode_predictions({k: v.cpu() for k, v in preds.items()})[0] gt_boxes = ds._boxes_in_ego(ds.sample_tokens[i]) p = os.path.join(args.out_dir, f"frame_{i:03d}.png") render_frame(s["images"].numpy(), gt_boxes, pred_boxes, p, args.score_thresh) paths.append(p) if args.gif and paths: from PIL import Image frames = [Image.open(p) for p in paths] gif_path = os.path.join(args.out_dir, "demo.gif") frames[0].save(gif_path, save_all=True, append_images=frames[1:], duration=700, loop=0) print(f"saved {gif_path}") if __name__ == "__main__": main()