Datasets:

syCen
/

action-worldmodel-bench

License:
action-worldmodel-bench / any_frame_test.py
Update any_frame_test.py
0ce4f63 verified 5 months ago
30.5 kB
"""
SAM 3 Interactive Video Annotator
==================================
A Gradio app for interactive video segmentation using SAM 3.
Usage:
    python sam3_annotator.py [--port 7860] [--share]
Then SSH tunnel:
    ssh -L 7860:holygpu8a11103:7860 scen@login.rc.fas.harvard.edu
Open http://localhost:7860 in your local browser.
"""

import argparse
import json
import os
import tempfile
import time
from pathlib import Path

import cv2
import gradio as gr
import numpy as np
import torch
from PIL import Image

from pycocotools import mask as mask_utils

# ── SAM 3 imports ──────────────────────────────────────────────────────────
from sam3.model_builder import build_sam3_video_predictor

# ── Global state ───────────────────────────────────────────────────────────
predictor = None  # initialized once in main


# ── Helpers ────────────────────────────────────────────────────────────────

def encode_mask_rle(mask: np.ndarray):
    mask = np.asfortranarray(mask.astype(np.uint8))
    rle = mask_utils.encode(mask)
    rle["counts"] = rle["counts"].decode("utf-8")
    return rle


def overlay_mask(frame, mask, color=(30, 144, 255), alpha=0.45):
    vis = frame.copy().astype(np.float32)
    color = np.array(color, dtype=np.float32)
    vis[mask] = vis[mask] * (1 - alpha) + color * alpha
    return vis.astype(np.uint8)

def extract_first_frame(video_path: str) -> np.ndarray:
    """Return the first frame of a video as RGB numpy array."""
    cap = cv2.VideoCapture(video_path)
    ret, frame = cap.read()
    cap.release()
    if not ret:
        raise ValueError(f"Cannot read video: {video_path}")
    return cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)


def get_video_info(video_path: str):
    """Return (width, height, fps, num_frames)."""
    cap = cv2.VideoCapture(video_path)
    w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
    h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
    fps = cap.get(cv2.CAP_PROP_FPS)
    n = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    cap.release()
    return w, h, fps, n


def load_all_frames(video_path: str):
    cap = cv2.VideoCapture(video_path)
    frames = []
    while True:
        ret, frame_bgr = cap.read()
        if not ret:
            break
        frames.append(cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB))
    cap.release()
    return frames


def masks_to_bboxes(masks: dict) -> list:
    """Convert a binary mask (H, W) → [x_min, y_min, x_max, y_max]."""
    bboxes = []
    for obj_id, mask in masks.items():
        if isinstance(mask, torch.Tensor):
            mask = mask.cpu().numpy()
        mask = mask.squeeze()
        if mask.ndim == 2:
            ys, xs = np.where(mask > 0)
            if len(xs) == 0:
                continue
            bboxes.append({
                "obj_id": int(obj_id),
                "bbox": [int(xs.min()), int(ys.min()), int(xs.max()), int(ys.max())],
            })
    return bboxes


def overlay_masks_on_frame(frame: np.ndarray, outputs: dict, alpha: float = 0.45) -> np.ndarray:
    """
    Overlay coloured segmentation masks on a frame.
    `outputs` is the raw dict returned by SAM 3 for one frame.
    Auto-detects key names since SAM3 output format may vary.
    """
    # Debug: print all keys so we can see what SAM3 returns
    print(f"[DEBUG overlay] outputs keys: {list(outputs.keys())}", flush=True)
    for k, v in outputs.items():
        if isinstance(v, torch.Tensor):
            print(f"[DEBUG overlay]   {k}: Tensor shape={v.shape}, dtype={v.dtype}", flush=True)
        elif isinstance(v, np.ndarray):
            print(f"[DEBUG overlay]   {k}: ndarray shape={v.shape}, dtype={v.dtype}", flush=True)
        elif isinstance(v, list):
            print(f"[DEBUG overlay]   {k}: list len={len(v)}, first type={type(v[0]) if v else 'empty'}", flush=True)
        else:
            print(f"[DEBUG overlay]   {k}: {type(v).__name__} = {v}", flush=True)

    vis = frame.copy().astype(np.float32)

    # Predefined distinct colours (RGB)
    COLORS = [
        (30, 144, 255),   # dodger blue
        (255, 80, 80),    # red
        (50, 205, 50),    # lime green
        (255, 165, 0),    # orange
        (180, 50, 255),   # purple
        (0, 255, 255),    # cyan
        (255, 255, 0),    # yellow
        (255, 105, 180),  # hot pink
    ]

    # ── Find obj_ids ──
    obj_ids = None
    for key in ["obj_ids", "out_obj_ids", "object_ids", "obj_id", "ids"]:
        if key in outputs and outputs[key] is not None:
            obj_ids = outputs[key]
            break
    if obj_ids is None:
        print("[DEBUG overlay] No obj_ids found, returning raw frame", flush=True)
        return frame

    # Convert to list if tensor
    if isinstance(obj_ids, torch.Tensor):
        obj_ids = obj_ids.cpu().tolist()
    elif isinstance(obj_ids, np.ndarray):
        obj_ids = obj_ids.tolist()

    # ── Find masks ──
    masks_raw = None
    for key in ["out_binary_masks", "masks", "video_res_masks", "low_res_masks", "pred_masks", "segmentations"]:
        if key in outputs and outputs[key] is not None:
            masks_raw = outputs[key]
            print(f"[DEBUG overlay] Using masks from key='{key}'", flush=True)
            break

    if masks_raw is None:
        print("[DEBUG overlay] No masks found, returning raw frame", flush=True)
        return frame

    # ── Convert masks to numpy bool array ──
    if isinstance(masks_raw, torch.Tensor):
        masks_np = (masks_raw > 0).cpu().numpy()
    elif isinstance(masks_raw, list):
        converted = []
        for m in masks_raw:
            if isinstance(m, torch.Tensor):
                converted.append((m > 0).cpu().numpy())
            else:
                converted.append(np.array(m) > 0)
        masks_np = np.array(converted)
    else:
        masks_np = np.array(masks_raw) > 0

    print(f"[DEBUG overlay] masks_np shape={masks_np.shape}, n_objects={len(obj_ids)}", flush=True)

    # ── Draw each mask ──
    vis_uint8 = vis.astype(np.uint8)  # for contour drawing
    for i, obj_id in enumerate(obj_ids):
        if i >= len(masks_np):
            break
        color = np.array(COLORS[i % len(COLORS)], dtype=np.float32)
        mask = masks_np[i].squeeze()  # collapse to (H, W)
        if mask.ndim != 2:
            print(f"[DEBUG overlay] Skipping obj {obj_id}: mask.ndim={mask.ndim} after squeeze", flush=True)
            continue

        n_pixels = mask.sum()
        print(f"[DEBUG overlay] obj_id={obj_id}: mask has {n_pixels} pixels", flush=True)

        if n_pixels == 0:
            continue

        # Blend colour onto float image
        vis[mask] = vis[mask] * (1 - alpha) + color * alpha

        # Draw contour on uint8 image
        contours, _ = cv2.findContours(
            mask.astype(np.uint8), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE
        )
        cv2.drawContours(vis_uint8, contours, -1, tuple(int(c) for c in color), 2)

    # Merge: use blended vis but paste contours on top
    result = vis.astype(np.uint8)
    # Where contours were drawn (vis_uint8 differs from original frame), use those pixels
    contour_mask = np.any(vis_uint8 != frame, axis=-1)
    result[contour_mask] = vis_uint8[contour_mask]

    return result


def draw_clicks_on_frame(frame: np.ndarray, clicks: list) -> np.ndarray:
    """Draw positive (green ★) and negative (red ★) clicks on the frame."""
    vis = frame.copy()
    for x, y, label in clicks:
        color = (0, 255, 0) if label == 1 else (255, 0, 0)
        # Draw star-like marker
        cv2.drawMarker(vis, (int(x), int(y)), color,
                       markerType=cv2.MARKER_STAR, markerSize=20, thickness=2)
    return vis


# ── Core workflow functions (called by Gradio) ─────────────────────────────

def upload_video(video_path_str):
    if not video_path_str or not video_path_str.strip():
        return None, "⚠️ Please enter a video path.", gr.update(interactive=False), gr.update(interactive=False), gr.update(interactive=False), None

    video_path = video_path_str.strip()
    if not os.path.exists(video_path):
        return None, f"⚠️ File not found: {video_path}", gr.update(interactive=False), gr.update(interactive=False), gr.update(interactive=False), None

    pred = predictor

    print(f"[VIDEO] Loading frames from {video_path} …", flush=True)
    t0 = time.time()
    frames = load_all_frames(video_path)
    if len(frames) == 0:
        raise ValueError(f"Cannot read video: {video_path}")

    first_frame = frames[0]
    w, h, fps, n_frames = get_video_info(video_path)
    print(f"[VIDEO] Loaded: {w}×{h}, {n_frames} frames — {time.time() - t0:.1f}s", flush=True)

    print(f"[VIDEO] Starting SAM3 session …", flush=True)
    t0 = time.time()
    response = pred.handle_request(
        request=dict(type="start_session", resource_path=video_path)
    )
    sid = response["session_id"]
    print(f"[VIDEO] Session started: {sid} — {time.time() - t0:.1f}s", flush=True)

    state = {
        "session_id": sid,
        "video_path": video_path,
        "width": w,
        "height": h,
        "fps": fps,
        "n_frames": n_frames,
        "frames": frames,
        "current_frame_idx": 0,
        "clicks_by_obj": {1: []},
        "current_obj_id": 1,
        "next_obj_id": 2,
        # "clicks": [],
        # "next_obj_id": 1,
        "current_outputs": None,
        "text_prompt": None,
    }

    info = f"✅ Video loaded: {w}×{h}, {fps:.1f} fps, {n_frames} frames"
    return (
        first_frame,
        info,
        gr.update(interactive=True),
        gr.update(interactive=True),
        gr.update(minimum=0, maximum=n_frames - 1, value=0, interactive=True),
        state,
    )


def select_frame(state, frame_idx):
    if state is None:
        return None, "⚠️ Please upload a video first.", state

    frame_idx = int(frame_idx)
    state["current_frame_idx"] = frame_idx
    # state["clicks"] = []
    state["clicks_by_obj"] = {}
    state["current_outputs"] = None
    state["text_prompt"] = None

    return state["frames"][frame_idx], f"📍 Selected frame {frame_idx}", state


def handle_click(state, evt: gr.SelectData, click_mode, obj_id):
    if state is None:
        return None, "⚠️ Please upload a video first.", state

    frame_idx = int(state.get("current_frame_idx", 0))
    obj_id = int(obj_id)

    x, y = evt.index
    label = 1 if click_mode == "Positive (include)" else 0

    if "clicks_by_obj" not in state:
        state["clicks_by_obj"] = {}

    state["clicks_by_obj"].setdefault(obj_id, [])
    state["clicks_by_obj"][obj_id].append((x, y, label))
    state["current_obj_id"] = obj_id
    state["text_prompt"] = None

    pred = predictor
    sid = state["session_id"]
    w, h = state["width"], state["height"]

    clicks = state["clicks_by_obj"][obj_id]
    points_abs = np.array([(cx, cy) for cx, cy, _ in clicks])
    labels = np.array([lb for _, _, lb in clicks])

    points_rel = torch.tensor(
        [[px / w, py / h] for px, py in points_abs],
        dtype=torch.float32,
    )
    labels_tensor = torch.tensor(labels, dtype=torch.int32)

    response = pred.handle_request(
        request=dict(
            type="add_prompt",
            session_id=sid,
            frame_index=frame_idx,
            points=points_rel,
            point_labels=labels_tensor,
            obj_id=obj_id,
        )
    )

    out = response["outputs"]
    state["current_outputs"] = out

    vis = overlay_masks_on_frame(state["frames"][frame_idx], out)

    # draw clicks for all objects
    for oid, clicks_i in state["clicks_by_obj"].items():
        vis = draw_clicks_on_frame(vis, clicks_i)

    total = sum(len(v) for v in state["clicks_by_obj"].values())
    info = f"🖱️ Frame {frame_idx}: object {obj_id}, total clicks={total}"

    return vis, info, state


# def handle_text_prompt(state, text_prompt):
#     """Alternative: use a text prompt instead of clicks."""
#     if state is None:
#         return None, "⚠️ Please upload a video first.", state
#     if not text_prompt.strip():
#         return None, "⚠️ Please enter a text prompt.", state

#     pred = predictor
#     sid = state["session_id"]

#     # Reset and clear clicks
#     pred.handle_request(request=dict(type="reset_session", session_id=sid))
#     state["clicks"] = []
#     state["text_prompt"] = text_prompt.strip()  # save for propagation

#     response = pred.handle_request(
#         request=dict(
#             type="add_prompt",
#             session_id=sid,
#             frame_index=prompt_frame_idx,
#             text=text_prompt.strip(),
#         )
#     )

#     out = response["outputs"]
#     state["current_outputs"] = out

#     vis = overlay_masks_on_frame(state["first_frame"], out)
#     info = f'📝 Text prompt: "{text_prompt.strip()}"'

#     return vis, info, state

def handle_text_prompt(state, text_prompt):
    return None, "⚠️ Text prompt is disabled for multi-object click mode.", state


def undo_last_click(state):
    if state is None:
        return None, "⚠️ No video loaded.", state

    frame_idx = int(state.get("current_frame_idx", 0))
    obj_id = int(state.get("current_obj_id", 1))
    curr_frame = state["frames"][frame_idx]

    clicks_by_obj = state.get("clicks_by_obj", {})
    clicks = clicks_by_obj.get(obj_id, [])

    if not clicks:
        return curr_frame.copy(), f"ℹ️ No clicks to undo for object {obj_id}.", state

    clicks.pop()

    pred = predictor
    sid = state["session_id"]
    w, h = state["width"], state["height"]

    pred.handle_request(request=dict(type="reset_session", session_id=sid))
    state["current_outputs"] = None

    vis = curr_frame.copy()

    for oid, obj_clicks in clicks_by_obj.items():
        if not obj_clicks:
            continue

        points_abs = np.array([(cx, cy) for cx, cy, _ in obj_clicks])
        labels = np.array([lb for _, _, lb in obj_clicks])
        points_rel = torch.tensor(
            [[px / w, py / h] for px, py in points_abs],
            dtype=torch.float32,
        )
        labels_tensor = torch.tensor(labels, dtype=torch.int32)

        response = pred.handle_request(
            request=dict(
                type="add_prompt",
                session_id=sid,
                frame_index=frame_idx,
                points=points_rel,
                point_labels=labels_tensor,
                obj_id=int(oid),
            )
        )

        state["current_outputs"] = response["outputs"]
        vis = overlay_masks_on_frame(curr_frame, response["outputs"])

    for _, obj_clicks in clicks_by_obj.items():
        vis = draw_clicks_on_frame(vis, obj_clicks)

    total = sum(len(v) for v in clicks_by_obj.values())
    return vis, f"↩️ Undone object {obj_id}. Total clicks={total}", state


def clear_all_clicks(state):
    if state is None:
        return None, "⚠️ No video loaded.", state

    pred = predictor
    sid = state["session_id"]
    pred.handle_request(request=dict(type="reset_session", session_id=sid))

    frame_idx = int(state.get("current_frame_idx", 0))
    state["clicks_by_obj"] = {}
    state["current_outputs"] = None
    state["text_prompt"] = None

    return state["frames"][frame_idx].copy(), "🗑️ All clicks cleared.", state


def propagate_and_export(state, progress=gr.Progress()):
    """Propagate masks, export bbox+mask JSON, and render overlay video."""
    if state is None:
        return None, None, "⚠️ No video loaded."
    if state["current_outputs"] is None:
        return None, None, "⚠️ No segmentation to propagate. Add clicks or text prompt first."

    pred = predictor
    sid = state["session_id"]
    w, h = state["width"], state["height"]
    n_frames = state["n_frames"]
    prompt_frame_idx = int(state.get("current_frame_idx", 0))

    # prompt_frame_idx = int(state.get("current_frame_idx", 0))

    print(
        f"[PROPAGATE] Using existing prompted state from frame {prompt_frame_idx} …",
        flush=True,
    )
    
    session = pred._get_session(sid)
    inference_state = session["state"]
    
    if inference_state["previous_stages_out"][prompt_frame_idx] is None:
        inference_state["previous_stages_out"][prompt_frame_idx] = "_THIS_FRAME_HAS_OUTPUTS_"

    progress(0, desc="Propagating masks through video…")

    all_frame_outputs = {}
    frame_count = 0
    for response in pred.handle_stream_request(
        # request=dict(type="propagate_in_video", session_id=sid)
        request=dict(
            type="propagate_in_video",
            session_id=sid,
            start_frame_index=prompt_frame_idx,
        )
    ):
        fidx = response["frame_index"]
        all_frame_outputs[fidx] = response["outputs"]
        frame_count += 1
        if frame_count % 10 == 0:
            progress(frame_count / n_frames, desc=f"Frame {frame_count}/{n_frames}")

    progress(0.9, desc="Exporting masks and rendering video…")

    result = {
        "video_path": state["video_path"],
        "width": w,
        "height": h,
        "fps": state["fps"],
        "n_frames": n_frames,
        "frames": {},
    }

    cap = cv2.VideoCapture(state["video_path"])
    raw_frames = []
    while True:
        ret, frame_bgr = cap.read()
        if not ret:
            break
        raw_frames.append(cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB))
    cap.release()

    overlay_frames = []

    for fidx in sorted(all_frame_outputs.keys()):
        out = all_frame_outputs[fidx]
        frame_entries = []

        frame_rgb = raw_frames[fidx].copy() if fidx < len(raw_frames) else np.zeros((h, w, 3), dtype=np.uint8)

        obj_ids = None
        for key in ["out_obj_ids", "obj_ids", "object_ids"]:
            if key in out and out[key] is not None:
                obj_ids = out[key]
                break

        if obj_ids is None:
            result["frames"][str(fidx)] = []
            overlay_frames.append(frame_rgb)
            continue

        if isinstance(obj_ids, torch.Tensor):
            obj_ids = obj_ids.cpu().numpy()
        obj_ids = np.asarray(obj_ids)

        masks_raw = None
        for key in ["out_binary_masks", "masks", "video_res_masks"]:
            if key in out and out[key] is not None:
                masks_raw = out[key]
                break

        masks_np = None
        if masks_raw is not None:
            if isinstance(masks_raw, torch.Tensor):
                masks_np = (masks_raw > 0).cpu().numpy()
            else:
                masks_np = np.asarray(masks_raw) > 0

        boxes_xywh = out.get("out_boxes_xywh", None)
        if boxes_xywh is not None:
            if isinstance(boxes_xywh, torch.Tensor):
                boxes_xywh = boxes_xywh.cpu().numpy()
            boxes_xywh = np.asarray(boxes_xywh)

        for i, obj_id in enumerate(obj_ids):
            entry = {"obj_id": int(obj_id)}

            mask = None
            if masks_np is not None and i < len(masks_np):
                mask = masks_np[i].squeeze()
                if mask.ndim == 2 and mask.sum() > 0:
                    ys, xs = np.where(mask > 0)
                    entry["bbox_xyxy"] = [int(xs.min()), int(ys.min()), int(xs.max()), int(ys.max())]
                    entry["mask_rle"] = encode_mask_rle(mask)
                    frame_rgb = overlay_mask(frame_rgb, mask)

            if "bbox_xyxy" not in entry and boxes_xywh is not None and i < len(boxes_xywh):
                bx, by, bw, bh = boxes_xywh[i]
                entry["bbox_xyxy"] = [
                    int(bx * w),
                    int(by * h),
                    int((bx + bw) * w),
                    int((by + bh) * h),
                ]

            frame_entries.append(entry)

        result["frames"][str(fidx)] = frame_entries
        overlay_frames.append(frame_rgb)

    out_dir = os.path.dirname(state["video_path"])
    json_path = os.path.join(out_dir, "sam3_bboxes_masks.json")
    with open(json_path, "w") as f:
        json.dump(result, f, indent=2)

    # overlay_video_path = os.path.join(tempfile.gettempdir(), "sam3_overlay.mp4")
    overlay_video_path = os.path.join(out_dir, "sam3_overlay.mp4")
    raw_overlay_path = os.path.join(out_dir, "sam3_overlay_raw.mp4")
    overlay_video_path = os.path.join(out_dir, "sam3_overlay.mp4")
    
    writer = cv2.VideoWriter(
        raw_overlay_path,  # 这里必须是 raw_overlay_path
        cv2.VideoWriter_fourcc(*"mp4v"),
        state["fps"],
        (w, h),
    )
    
    for frame_rgb in overlay_frames:
        frame_bgr = cv2.cvtColor(frame_rgb, cv2.COLOR_RGB2BGR)
        writer.write(frame_bgr)
    
    writer.release()
    
    # Convert to browser-compatible H264
    ret = os.system(
        f'ffmpeg -y -i "{raw_overlay_path}" '
        f'-c:v libx264 -pix_fmt yuv420p -movflags +faststart '
        f'"{overlay_video_path}"'
    )
    
    if ret != 0:
        print("[WARN] ffmpeg failed, falling back to raw mp4", flush=True)
        overlay_video_path = raw_overlay_path
    
    progress(1.0, desc="Done!")
    info = f"✅ Done. JSON saved to: {json_path}\n🎬 Video saved to: {overlay_video_path}"
    return json_path, overlay_video_path, info


# ── Gradio UI ──────────────────────────────────────────────────────────────

def build_app():
    with gr.Blocks(
        title="SAM 3 Video Annotator",
    ) as app:
        gr.Markdown("# 🎯 SAM 3 — Interactive Video Annotator", elem_classes="main-title")
        gr.Markdown(
            "Enter a video path on the server → click on the first frame to select objects → "
            "propagate through the entire video → download bounding boxes as JSON."
        )

        state = gr.State(None)

        with gr.Row():
            # ── Left panel: controls ──
            with gr.Column(scale=1):
                # video_input = gr.Textbox(
                #     label="📹 Video Path (on server)",
                #     placeholder="/path/to/video.mp4 or /path/to/jpeg_frames_dir/",
                #     info="Enter the absolute path to an MP4 file or a JPEG frames directory on the server.",
                # )
                DEFAULT_VIDEO_PATH = "/net/holy-isilon/ifs/rc_labs/ydu_lab/sycen/data/omnivitac/grasping_partial/mnt/oss_data/it20260119/Grasping/grasping_gello_flip/videos/0000/camera2.mp4"

                video_input = gr.Textbox(
                    label="📹 Video Path (on server)",
                    value=DEFAULT_VIDEO_PATH,
                    info="Enter the absolute path to an MP4 file or a JPEG frames directory on the server.",
                )
                upload_btn = gr.Button("🚀 Load Video", variant="primary")

                gr.Markdown("---")
                gr.Markdown("### Click Prompts")

                obj_id_input = gr.Number(
                    label="Object ID",
                    value=1,
                    precision=0,
                )
                
                click_mode = gr.Radio(
                    choices=["Positive (include)", "Negative (exclude)"],
                    value="Positive (include)",
                    label="Click Mode",
                )
                with gr.Row():
                    undo_btn = gr.Button("↩️ Undo", interactive=False)
                    clear_btn = gr.Button("🗑️ Clear All", interactive=False)

                gr.Markdown("---")
                gr.Markdown("### Text Prompt (alternative)")
                text_input = gr.Textbox(
                    label="Text prompt",
                    placeholder='e.g. "person", "red car", "dog"',
                )
                text_btn = gr.Button("📝 Apply Text Prompt")

                gr.Markdown("---")
                propagate_btn = gr.Button(
                    "▶️ Propagate & Export BBoxes", variant="primary", interactive=False
                )

            # ── Right panel: display ──
            with gr.Column(scale=2):
                # frame_display = gr.Image(
                #     label="First Frame (click to annotate)",
                #     interactive=False,
                #     type="numpy",
                # )

                # video_preview = gr.Video(
                #     label="Original Video",
                #     value=DEFAULT_VIDEO_PATH,
                # )
                
                frame_slider = gr.Slider(
                    minimum=0,
                    maximum=1,
                    step=1,
                    value=0,
                    label="Annotation frame index",
                    interactive=False,
                )
                
                frame_display = gr.Image(
                    label="Selected Frame (click to annotate)",
                    interactive=True,
                    type="numpy",
                )
                
                status = gr.Textbox(
                    label="Status",
                    interactive=False,
                    elem_classes="status-box",
                    value="👆 Upload a video to get started.",
                )
                json_output = gr.File(label="📦 Download BBox JSON", visible=True)
                video_output = gr.Video(label="🎬 Propagated Tracking Video")

        # ── Wiring ─────────────────────────────────────────────────────────
        upload_btn.click(
            fn=upload_video,
            inputs=[video_input],
            outputs=[frame_display, status, undo_btn, propagate_btn, frame_slider, state],
        ).then(
            fn=lambda: gr.update(interactive=True),
            outputs=[clear_btn],
        )
        
        frame_slider.change(
            fn=select_frame,
            inputs=[state, frame_slider],
            outputs=[frame_display, status, state],
        )

        # Click on image
        frame_display.select(
            fn=handle_click,
            inputs=[state, click_mode, obj_id_input],
            outputs=[frame_display, status, state],
        )

        undo_btn.click(
            fn=undo_last_click,
            inputs=[state],
            outputs=[frame_display, status, state],
        )

        clear_btn.click(
            fn=clear_all_clicks,
            inputs=[state],
            outputs=[frame_display, status, state],
        )

        text_btn.click(
            fn=handle_text_prompt,
            inputs=[state, text_input],
            outputs=[frame_display, status, state],
        )

        propagate_btn.click(
            fn=propagate_and_export,
            inputs=[state],
            outputs=[json_output, video_output, status],
        )

    return app


# ── Main ───────────────────────────────────────────────────────────────────

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="SAM 3 Interactive Video Annotator")
    parser.add_argument("--port", type=int, default=7860)
    parser.add_argument("--share", action="store_true", help="Create a public Gradio link")
    parser.add_argument("--host", type=str, default="0.0.0.0")
    args = parser.parse_args()

    # ── Load model ONCE at startup (before any requests) ──
    import sys
    t_total = time.time()

    print("=" * 60, flush=True)
    print("[INIT] Step 1/4: Importing sam3 internals …", flush=True)
    t0 = time.time()
    from sam3.model_builder import (
        build_sam3_video_predictor as _build_pred,
    )
    print(f"[INIT] Step 1/4 done. Import took {time.time() - t0:.1f}s", flush=True)

    print("[INIT] Step 2/4: Calling build_sam3_video_predictor() …", flush=True)
    print("[INIT]   (this downloads/loads checkpoints — may be slow on /net storage)", flush=True)
    t0 = time.time()

    # Monkey-patch torch.load to add timing
    _original_torch_load = torch.load
    def _timed_torch_load(*args, **kwargs):
        path_str = str(args[0]) if args else str(kwargs.get("f", "???"))
        # Truncate long paths
        display = path_str if len(path_str) < 100 else "…" + path_str[-80:]
        print(f"[INIT]     torch.load: {display}", flush=True)
        t = time.time()
        result = _original_torch_load(*args, **kwargs)
        print(f"[INIT]     torch.load done — {time.time() - t:.1f}s", flush=True)
        return result
    torch.load = _timed_torch_load

    predictor = _build_pred()

    # Restore original
    torch.load = _original_torch_load
    print(f"[INIT] Step 2/4 done. Model build took {time.time() - t0:.1f}s", flush=True)

    print("[INIT] Step 3/4: Moving model to GPU / compiling …", flush=True)
    t0 = time.time()
    # Force a sync to make sure everything is on GPU
    if torch.cuda.is_available():
        torch.cuda.synchronize()
        mem = torch.cuda.memory_allocated() / 1024**3
        print(f"[INIT]   GPU memory used: {mem:.2f} GB", flush=True)
    print(f"[INIT] Step 3/4 done. {time.time() - t0:.1f}s", flush=True)

    print(f"[INIT] Step 4/4: Building Gradio app …", flush=True)
    t0 = time.time()
    app = build_app()
    print(f"[INIT] Step 4/4 done. {time.time() - t0:.1f}s", flush=True)

    print("=" * 60, flush=True)
    print(f"[INIT] Total startup: {time.time() - t_total:.1f}s", flush=True)
    print(f"[INIT] Launching server on {args.host}:{args.port}", flush=True)
    print("=" * 60, flush=True)

    app.launch(
        allowed_paths=[
            "/net/holy-isilon/ifs/rc_labs/ydu_lab/sycen/data/omnivitac/grasping_partial/mnt/oss_data/it20260119/Grasping/grasping_gello_flip/videos/0000"
        ],
        server_name=args.host,
        server_port=args.port,
        share=args.share,
        theme=gr.themes.Soft(),
        css=".main-title { text-align: center; } .status-box { font-family: monospace; font-size: 0.9em; }",
    )