detect_scene.py

import torch
import numpy as np
from pathlib import Path
from transnetv2_pytorch import TransNetV2
import json
import re

# SCENE_CUT_THRESHOLD = 0.09
K = 3 # Number of cuts to detect
MIN_DURATION_FRAMES = 2
MIN_CONFIDENCE = 0.02

data_dir = Path("data/animations")
files = sorted(data_dir.glob("sample-*.webp"))
print(f"Found {len(files)} files to process.")
    
def get_best_device():
    if torch.cuda.is_available():
        return torch.device("cuda")
    elif torch.backends.mps.is_available():
        return torch.device("mps")
        # return torch.device("cpu")
    else:
        return torch.device("cpu")

def load_original_frames(filepath):
    """Load original frames from an animated webp file as PIL Images."""
    from PIL import Image
    im = Image.open(filepath)
    frames = []
    try:
        while True:
            frames.append(im.convert("RGB"))
            im.seek(im.tell() + 1)
    except EOFError:
        pass
    return frames

def save_prediction_plot(single_frame_pred, original_frames, filename, interval=5, title=None):
    """
    Save a plot of single frame predictions with thumbnails annotated at regular intervals.
    """
    import matplotlib.pyplot as plt
    from matplotlib.offsetbox import OffsetImage, AnnotationBbox
    import numpy as np

    plt.figure(figsize=(12, 4))
    if title:
        plt.title(title)
    plt.plot(single_frame_pred)
    ax = plt.gca()
    # Add thumbnails at regular intervals
    for idx in range(0, len(original_frames), interval):
        thumb = original_frames[idx].resize((64, 64))
        imagebox = OffsetImage(np.array(thumb), zoom=0.5)
        ab = AnnotationBbox(imagebox, (idx, single_frame_pred[idx]), frameon=False, box_alignment=(0.5, -0.1))
        ax.add_artist(ab)
    plt.xlabel("Frame")
    plt.ylabel("Prediction")
    plt.tight_layout()
    plt.savefig(filename)
    plt.close()

def save_timeline_jpg(frames, scene_change_indices, filename, interval=5, roi_radius=2, title=None, single_frame_pred=None):
    """
    Save a timeline JPG with thumbnails every `interval` frames and every frame near scene changes.
    Scene change regions are highlighted. Each thumbnail is annotated with its frame index.
    """
    import matplotlib.pyplot as plt
    from matplotlib.offsetbox import OffsetImage, AnnotationBbox
    import matplotlib.patches as mpatches
    import numpy as np

    # Determine frames to render
    frames_to_render = set(range(0, len(frames), interval))
    for idx in scene_change_indices:
        for offset in range(-roi_radius, roi_radius+1):
            fidx = idx + offset
            if 0 <= fidx < len(frames):
                frames_to_render.add(fidx)
    frames_to_render = sorted(frames_to_render)

    # Map frames to evenly spaced positions
    n = len(frames_to_render)
    x_positions = list(range(n))

    fig, ax = plt.subplots(figsize=(max(8, n*0.5), 3))
    ax.set_xlim(-1, n)
    ax.set_ylim(0, 1)
    ax.axis('off')

    # Highlight scene change regions
    for idx in scene_change_indices:
        region = [i for i, fidx in enumerate(frames_to_render) if abs(fidx-idx) <= roi_radius]
        if region:
            start, end = region[0], region[-1]
            rect = mpatches.Rectangle((start-0.5, 0.05), end-start+1, 0.9, color='yellow', alpha=0.2)
            ax.add_patch(rect)

    # Prepare sets for quick lookup
    last_frames = set(scene_change_indices)
    first_frames = set(idx + 1 for idx in scene_change_indices if idx + 1 < len(frames))

    # Draw thumbnails and annotate
    for i, fidx in enumerate(frames_to_render):
        thumb = frames[fidx].resize((32, 32))
        imagebox = OffsetImage(np.array(thumb), zoom=0.7)
        # Determine border color
        if fidx in last_frames:
            bboxprops = dict(edgecolor='red', linewidth=2)
        elif fidx in first_frames:
            bboxprops = dict(edgecolor='green', linewidth=2)
        else:
            bboxprops = None
        ab = AnnotationBbox(
            imagebox,
            (x_positions[i], 0.6),
            frameon=True,
            box_alignment=(0.5, 0.5),
            bboxprops=bboxprops
        )
        ax.add_artist(ab)
        # Draw frame index
        ax.text(x_positions[i], 0.32, str(fidx), ha='center', va='center', fontsize=9, color='black', bbox=dict(facecolor='white', edgecolor='none', alpha=0.8, boxstyle='round,pad=0.2'))
        # Draw prediction value below frame index
        if single_frame_pred is not None:
            pred_val = single_frame_pred[fidx]
            # Ensure pred_val is a scalar float for formatting
            if isinstance(pred_val, np.ndarray):
                pred_val = float(pred_val.squeeze())
            ax.text(x_positions[i], 0.18, f"{pred_val:.2f}", ha='center', va='center', fontsize=8, color='blue', bbox=dict(facecolor='white', edgecolor='none', alpha=0.7, boxstyle='round,pad=0.2'))

    if title:
        ax.text(0, 0.95, title, fontsize=12, ha='left', va='top', color='navy')
    plt.tight_layout()
    plt.savefig(filename, dpi=150)
    plt.close(fig)

def frames_to_video_tensor(frames):
    """Convert a list of PIL frames to a torch tensor of shape (num_frames, 27, 48, 3) and dtype uint8."""
    import numpy as np
    from PIL import Image
    processed = []
    for frame in frames:
        arr = np.array(frame.resize((48, 27), resample=Image.Resampling.BILINEAR), dtype=np.uint8)
        processed.append(torch.from_numpy(arr))
    return torch.stack(processed)

def detect_scene_changes(frames):
    video_tensor = frames_to_video_tensor(frames)
    video_tensor = video_tensor.unsqueeze(0).to(device)  # shape: 1 x num_frames x H x W x 3
    with torch.no_grad():
        single_frame_logits, all_frame_logits = model(video_tensor)
        # Squeeze last dimension so output is flat (num_frames,)
        single_frame_logits_np = single_frame_logits.cpu().numpy().squeeze()  # shape: (num_frames,)
        all_frame_logits_np = all_frame_logits["many_hot"].cpu().numpy().squeeze()  # shape: (num_frames,)
        single_frame_pred = torch.sigmoid(single_frame_logits).cpu().numpy().squeeze()  # shape: (num_frames,)
        all_frame_pred_np = torch.sigmoid(all_frame_logits["many_hot"]).cpu().numpy().squeeze()  # shape: (num_frames,)
    return {
        "single_frame_pred": single_frame_pred,
        "all_frame_pred": all_frame_pred_np,
        "single_frame_logits": single_frame_logits_np,
        "all_frame_logits": all_frame_logits_np,
    }

def cached_detect_scene_changes(file, original_frames):
    """Detect scene changes with caching to avoid redundant computation."""
    match = re.search(r"sample-(\d+)", file.name)
    sample_num = match.group(1) if match else "unknown"
    transnetv2_json = file.parent / f"sample-{sample_num}.transnetv2.json"
    if transnetv2_json.exists():
        with open(transnetv2_json, "r") as f:
            result = json.load(f)
        result["single_frame_pred"] = np.array(result["single_frame_pred"])
        result["all_frame_pred"] = np.array(result["all_frame_pred"])
        result["single_frame_logits"] = np.array(result["single_frame_logits"])
        result["all_frame_logits"] = np.array(result["all_frame_logits"])
    else:
        result = detect_scene_changes(original_frames)
        # Save model output to cache file
        with open(transnetv2_json, "w") as f:
            json.dump({
                "single_frame_pred": result["single_frame_pred"].tolist(),
                "all_frame_pred": result["all_frame_pred"].tolist(),
                "single_frame_logits": result["single_frame_logits"].tolist(),
                "all_frame_logits": result["all_frame_logits"].tolist()
            }, f, indent=2)
    return result

if __name__ == "__main__":
    device = get_best_device()
    print(f"Using device: {device}")
    model = TransNetV2()
    state_dict = torch.load("transnetv2-pytorch-weights.pth")
    model.load_state_dict(state_dict)
    model.eval().to(device)

    for file in files:
        match = re.search(r"sample-(\d+)", file.name)
        sample_num = match.group(1) if match else "unknown"
        original_frames = load_original_frames(file)
        result = cached_detect_scene_changes(file, original_frames)
        # scene_change_indices = [i for i, p in enumerate(result["single_frame_pred"]) if p >= SCENE_CUT_THRESHOLD]
        # Detect top-K-1 scene changes
        single_frame_pred = result["single_frame_pred"]
        
        # Ignore first and last frame when selecting scene changes, and enforce MIN_DURATION_FRAMES between cuts
        valid_indices = np.arange(1, len(single_frame_pred) - 1)
        valid_preds = single_frame_pred[1:-1]
        # Sort indices by prediction score (descending)
        sorted_indices = valid_indices[np.argsort(valid_preds)[::-1]]
        scene_change_indices = []
        scene_cut_confidences = []
        for idx in sorted_indices:
            if all(abs(idx - prev) >= MIN_DURATION_FRAMES for prev in scene_change_indices):
                scene_change_indices.append(int(idx))
                scene_cut_confidences.append(float(single_frame_pred[idx]))
            if len(scene_change_indices) >= (K - 1):
                break

        # Check if any confidence is below MIN_CONFIDENCE
        failed = any(conf < MIN_CONFIDENCE for conf in scene_cut_confidences)

        print(f"File: {file.name}, Frames: {len(original_frames)}, Scene Changes: {len(scene_change_indices)}, Success: {not failed}")
        # Save results to JSON (include threshold and predictions)
        json_filename = file.parent / f"sample-{sample_num}.json"
        with open(json_filename, "w") as f:
            json.dump({
                "num_frames": len(original_frames),
                "scene_change_indices": scene_change_indices,
                "scene_cut_confidences": scene_cut_confidences,
                # "threshold": SCENE_CUT_THRESHOLD
                "params": {
                    "k": K,
                    "min_duration_frames": MIN_DURATION_FRAMES,
                    "min_confidence": MIN_CONFIDENCE,
                },
                "success": not failed
            }, f, indent=2)
        # Save timeline JPG
        timeline_filename = file.parent / f"sample-{sample_num}.timeline.jpg"
        plot_filename = file.parent / f"sample-{sample_num}.plot.jpg"
        save_timeline_jpg(
            frames=original_frames,
            scene_change_indices=scene_change_indices,
            filename=timeline_filename,
            interval=10,
            roi_radius=2,
            title=f"Timeline: {file.name}",
            single_frame_pred=result["single_frame_pred"]
        )
        # Save prediction plot with thumbnails
        save_prediction_plot(
            single_frame_pred=result["single_frame_pred"],
            original_frames=original_frames,
            filename=plot_filename,
            interval=5,
            title=f"Single Frame Predictions: {file.name}"
        )