tools/render_demo.py

"""Render a multi-view demo video of the 3D point tracks: all views side by side, persistent points with
fading trails, size and brightness from camera depth. No text on the frames.

The clips given on the command line are played back to back (consecutive subtask clips of one episode
tile it, so the point identities carry over).

usage: python render_demo.py --out demo.mp4 [--height 384] <clip_dir> [<clip_dir> ...]
"""
import argparse
import glob
import json
import os

import cv2
import numpy as np

SS = 3  # supersampling of the drawing layer


def view_order(name):
    return (0, int(name[4:])) if name.startswith("view") and name[4:].isdigit() else (1, name)


def read_video(path):
    import av

    with av.open(path) as c:
        return np.stack([f.to_ndarray(format="rgb24") for f in c.decode(video=0)])


def load_clip(d):
    meta = json.load(open(os.path.join(d, "meta.json")))
    n = meta["n_frames"]
    names = sorted(meta["views"], key=view_order)
    views = {}
    for v in names:
        m = meta["views"][v]
        intr = m["intrinsics"]
        T = np.asarray(m.get("T_cam_world_perframe", m["T_cam_world"]), np.float64)
        if T.ndim == 2:
            T = np.repeat(T[None], n, 0)
        views[v] = dict(
            K=np.array([[intr["fx"], 0, intr["cx"]], [0, intr["fy"], intr["cy"]], [0, 0, 1]], np.float64),
            dist=np.array([intr.get(k, 0.0) for k in ("k1", "k2", "p1", "p2", "k3")], np.float64),
            wh=(intr["width"], intr["height"]), T=T[:n], rgb=read_video(os.path.join(d, f"{v}.mp4"))[:n])
    parts = {}
    for f in ["robot.npz", "gripper.npz"] + sorted(os.path.basename(p) for p in glob.glob(os.path.join(d, "object*.npz"))):
        if os.path.isfile(os.path.join(d, f)):
            parts[f[:-4]] = np.load(os.path.join(d, f))["points"].astype(np.float64)[:n]
    # view1 camera frame -> world, per frame
    Tw = np.linalg.inv(views["view1"]["T"])
    world = {k: np.einsum("tij,tnj->tni", Tw[:, :3, :3], p) + Tw[:, None, :3, 3] for k, p in parts.items()}
    return meta, names, views, world


def concat_clips(dirs):
    clips = [load_clip(d) for d in dirs]
    names = clips[0][1]
    views = {v: {k: clips[0][2][v][k] for k in ("K", "dist", "wh")} for v in names}
    for v in names:
        views[v]["T"] = np.concatenate([c[2][v]["T"] for c in clips])
        views[v]["rgb"] = np.concatenate([c[2][v]["rgb"] for c in clips])
    keys = list(clips[0][3])
    world = {k: np.concatenate([c[3][k] for c in clips]) for k in keys if all(k in c[3] for c in clips)}
    return clips[0][0], names, views, world


def farthest(X, k):
    if k >= len(X):
        return np.arange(len(X))
    sel = [int(np.argmin(np.linalg.norm(X - X.mean(0), axis=1)))]
    dmin = np.linalg.norm(X - X[sel[0]], axis=1)
    for _ in range(k - 1):
        sel.append(int(np.argmax(dmin)))
        dmin = np.minimum(dmin, np.linalg.norm(X - X[sel[-1]], axis=1))
    return np.sort(np.asarray(sel))


def gradient_colors(X, cmap, lo, hi):
    import matplotlib

    Xc = X - X.mean(0)
    _, _, Vt = np.linalg.svd(Xc, full_matrices=False)
    p = Xc @ Vt[:2].T
    a = (p[:, 0] - p[:, 0].min()) / max(np.ptp(p[:, 0]), 1e-9)
    b = (p[:, 1] - p[:, 1].min()) / max(np.ptp(p[:, 1]), 1e-9)
    rgb = matplotlib.colormaps[cmap](lo + (hi - lo) * a)[:, :3] * (0.85 + 0.15 * b)[:, None]
    return np.clip(rgb, 0, 1)


OBJECT_CMAPS = [("spring", 0.0, 0.55), ("cool", 0.35, 1.0), ("Wistia", 0.2, 1.0), ("PuRd", 0.45, 0.95)]


def select_points(world, targets, n_body, n_object):
    """Farthest-point subsampling on the frame-0 shape: arm and gripper together (even spacing over the
    whole robot), then each object; fixed colour per point."""
    pts, cols, is_obj = [], [], []
    body = [world[k] for k in ("robot", "gripper") if k in world]
    if body:
        B = np.concatenate(body, 1)
        B = B[:, farthest(B[0], n_body)]
        pts.append(B)
        cols.append(gradient_colors(B[0], "turbo", 0.08, 0.92))
        is_obj.append(np.zeros(B.shape[1], bool))
    objs = []
    for k in sorted(k for k in world if k.startswith("object")):
        O = world[k]
        sizes = targets.get(k)
        if sizes and sum(sizes) == O.shape[1] and len(sizes) > 1:  # several targets packed in one file
            objs += np.split(O, np.cumsum(sizes)[:-1], axis=1)
        else:
            objs.append(O)
    for i, O in enumerate(objs):
        O = O[:, farthest(O[0], n_object)]
        cm, lo, hi = OBJECT_CMAPS[i % len(OBJECT_CMAPS)]
        pts.append(O)
        cols.append(gradient_colors(O[0], cm, lo, hi))
        is_obj.append(np.ones(O.shape[1], bool))
    return np.concatenate(pts, 1), (np.concatenate(cols) * 255).astype(np.uint8), np.concatenate(is_obj)


def project(view, Xw, scale):
    """[T,N,3] world -> uv [T,N,2] output pixels, depth [T,N], valid [T,N]."""
    T = view["T"]
    Xc = np.einsum("tij,tnj->tni", T[:, :3, :3], Xw) + T[:, None, :3, 3]
    z = Xc[..., 2]
    w, h = view["wh"]
    K = view["K"]
    uv = np.zeros(Xc.shape[:2] + (2,))
    ok = z > 0.05
    x = Xc[..., 0] / np.where(ok, z, 1)
    y = Xc[..., 1] / np.where(ok, z, 1)
    u0, v0 = K[0, 0] * x + K[0, 2], K[1, 1] * y + K[1, 2]
    ok &= (u0 > -0.1 * w) & (u0 < 1.1 * w) & (v0 > -0.1 * h) & (v0 < 1.1 * h)  # trust distortion only near the image
    if np.any(view["dist"]):
        sel = np.nonzero(ok)
        p, _ = cv2.projectPoints(Xc[sel].reshape(-1, 1, 3), np.zeros(3), np.zeros(3), K, view["dist"])
        uv[sel] = p.reshape(-1, 2)
    else:
        uv[..., 0], uv[..., 1] = u0, v0
    uv = (uv + 0.5) * scale - 0.5
    return uv, z, ok


def draw(frame, uv, z, ok, cols, is_obj, t, tail, base_r):
    H, W = frame.shape[:2]
    h, w = H * SS, W * SS
    rgb = np.zeros((h, w, 3), np.uint8)
    a = np.zeros((h, w), np.uint8)
    P = uv * SS
    inside = ok & (uv[..., 0] >= -2) & (uv[..., 0] < W + 2) & (uv[..., 1] >= -2) & (uv[..., 1] < H + 2)
    zt = z[t][inside[t]]
    zref = np.median(zt) if len(zt) else 1.0
    rad = np.clip(base_r * zref / np.maximum(z, 1e-3), 0.65 * base_r, 1.6 * base_r)
    rad = np.where(is_obj[None], rad * 1.15, rad)
    near = np.clip((zref - z) / max(0.25 * zref, 1e-3), -1, 1)  # >0 nearer than the median
    shade = 0.82 + 0.18 * near
    # trails: oldest segment first
    for k in range(tail, 0, -1):
        t0, t1 = t - k, t - k + 1
        if t0 < 0:
            continue
        f = 1 - (k - 1) / tail
        al = int(255 * (0.05 + 0.55 * f ** 1.5))
        seg = np.nonzero(inside[t0] & inside[t1])[0]
        for n in seg:
            th = max(1, int(round(SS * rad[t1, n] * (0.25 + 0.45 * f))))
            p0 = tuple(int(round(c * 16)) for c in P[t0, n])
            p1 = tuple(int(round(c * 16)) for c in P[t1, n])
            c = tuple(int(x) for x in cols[n] * shade[t1, n])
            cv2.line(rgb, p0, p1, c, th, cv2.LINE_AA, shift=4)
            cv2.line(a, p0, p1, al, th, cv2.LINE_AA, shift=4)
    # dots: far first, near on top; rim (dark on the robot, white on objects), colour body, highlight
    order = np.argsort(-z[t])
    for n in order:
        if not inside[t, n]:
            continue
        cen = tuple(int(round(c * 16)) for c in P[t, n])
        r = rad[t, n] * SS
        col = cols[n].astype(np.float32) * shade[t, n]
        rim = (255, 255, 255) if is_obj[n] else tuple(int(x) for x in col * 0.35)
        cv2.circle(rgb, cen, int(round((r + 0.9 * SS) * 16)), rim, -1, cv2.LINE_AA, shift=4)
        cv2.circle(a, cen, int(round((r + 0.9 * SS) * 16)), 235, -1, cv2.LINE_AA, shift=4)
        cv2.circle(rgb, cen, int(round(r * 16)), tuple(int(x) for x in col), -1, cv2.LINE_AA, shift=4)
        hl = tuple(int(round((c - 0.3 * r) * 16)) for c in P[t, n])
        hcol = tuple(int(min(255, x * 0.55 + 255 * 0.45)) for x in col)
        cv2.circle(rgb, hl, int(round(0.38 * r * 16)), hcol, -1, cv2.LINE_AA, shift=4)
    af = a.astype(np.float32) / 255.0
    pre = rgb.astype(np.float32) / 255.0 * af[..., None]
    pre = cv2.resize(pre, (W, H), interpolation=cv2.INTER_AREA)
    af = cv2.resize(af, (W, H), interpolation=cv2.INTER_AREA)
    return np.clip(frame.astype(np.float32) * (1 - af[..., None]) + pre * 255.0, 0, 255).round().astype(np.uint8)


class Writer:
    def __init__(self, path, w, h, fps, crf):
        import av

        self.c = av.open(path, "w", options={"movflags": "+faststart"})
        self.s = self.c.add_stream("libx264", rate=int(fps))
        self.s.width, self.s.height, self.s.pix_fmt = w, h, "yuv420p"
        self.s.options = {"crf": str(crf), "preset": "slow", "threads": "8"}

    def write(self, frame):
        import av

        for p in self.s.encode(av.VideoFrame.from_ndarray(np.ascontiguousarray(frame), format="rgb24")):
            self.c.mux(p)

    def close(self):
        for p in self.s.encode():
            self.c.mux(p)
        self.c.close()


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("clips", nargs="+")
    ap.add_argument("--out", required=True)
    ap.add_argument("--height", type=int, default=384, help="height of each view in the output")
    ap.add_argument("--n-body", type=int, default=128, help="points on arm + gripper")
    ap.add_argument("--n-object", type=int, default=40)
    ap.add_argument("--tail", type=int, default=10)
    ap.add_argument("--radius", type=float, default=0.011, help="dot radius as a fraction of the view height")
    ap.add_argument("--views", default=None, help="comma-separated subset/order of views")
    ap.add_argument("--gap", type=int, default=4, help="white gap between views (px)")
    ap.add_argument("--crf", type=int, default=20)
    ap.add_argument("--poster", default=None, help="PNG of frame --poster-frame")
    ap.add_argument("--poster-frame", type=float, default=0.5, help="fraction of the video")
    args = ap.parse_args()

    meta, names, views, world = concat_clips(args.clips)
    if args.views:
        names = args.views.split(",")
    obj = meta.get("object") if isinstance(meta.get("object"), dict) else {}
    targets = {"object": [t["n_points"] for t in obj.get("targets", []) if "n_points" in t]}
    Xw, cols, is_obj = select_points(world, targets, args.n_body, args.n_object)
    n_t = Xw.shape[0]
    tiles = []
    for v in names:
        vw = views[v]
        h0, w0 = vw["rgb"].shape[1:3]
        s = args.height / h0
        W = int(round(w0 * s / 2)) * 2
        uv, z, ok = project(vw, Xw, np.array([W / w0, args.height / h0]))
        if (w0, h0) != tuple(vw["wh"]):
            raise ValueError(f"{v}: video {w0}x{h0} != intrinsics {vw['wh']}")
        T = vw["T"]
        centers = -np.einsum("tji,tj->ti", T[:, :3, :3], T[:, :3, 3])
        moving = np.linalg.norm(centers - centers[0], axis=1).max() > 0.01
        tiles.append(dict(v=v, W=W, uv=uv, z=z, ok=ok, tail=0 if moving else args.tail))
    width = sum(t["W"] for t in tiles) + args.gap * (len(tiles) - 1)
    width += width % 2
    fps = meta.get("fps", 15)
    wr = Writer(args.out, width, args.height, fps, args.crf)
    poster_t = int(args.poster_frame * (n_t - 1))
    for t in range(n_t):
        row = []
        for i, tl in enumerate(tiles):
            fr = cv2.resize(views[tl["v"]]["rgb"][t], (tl["W"], args.height), interpolation=cv2.INTER_LANCZOS4
                            if tl["W"] > views[tl["v"]]["rgb"].shape[2] else cv2.INTER_AREA)
            row.append(draw(fr, tl["uv"], tl["z"], tl["ok"], cols, is_obj, t, tl["tail"], args.radius * args.height))
            if i < len(tiles) - 1:
                row.append(np.full((args.height, args.gap, 3), 255, np.uint8))
        canvas = np.concatenate(row, 1)
        if canvas.shape[1] < width:
            canvas = np.pad(canvas, ((0, 0), (0, width - canvas.shape[1]), (0, 0)), constant_values=255)
        wr.write(canvas)
        if args.poster and t == poster_t:
            cv2.imwrite(args.poster, canvas[..., ::-1])
    wr.close()
    print(f"{args.out}: {len(tiles)} views, {Xw.shape[1]} points ({int(is_obj.sum())} on objects), {n_t} frames @ {fps} fps, {width}x{args.height}")


if __name__ == "__main__":
    main()