droid/scripts/visualize_50_episodes_mesh.py

#!/usr/bin/env python3
"""
Visualize mesh GT overlays for 50 episodes to identify alignment issues.
Simply shows mesh vertices overlaid on video frames for visual inspection.
"""
import sys
from pathlib import Path
sys.path.append(str(Path(__file__).parent.parent))

import numpy as np
import cv2
import mediapy as media
from tqdm import tqdm

# Load episodes
data_path = Path('/mnt/kevin/data/droid_preprocessed/data')
episode_files = sorted(list(data_path.glob("episode_*.npz")))

# Use first 50 episodes
split_idx = int(0.9 * len(episode_files))
train_episodes = episode_files[:split_idx]
samples_to_viz = train_episodes[:50]

output_dir = Path('mesh_overlay_50_episodes')
output_dir.mkdir(exist_ok=True)

print(f"Visualizing {len(samples_to_viz)} episodes with mesh overlays")
print(f"Saving individual videos to {output_dir}")

for ep_idx, npz_file in enumerate(tqdm(samples_to_viz, desc="Processing")):
    data = np.load(npz_file)

    # Load data
    images_ext = data['images_exterior']  # [T, 180, 320, 3]
    images_wrist = data['images_wrist']  # [T, 180, 320, 3]

    mesh_ext = data['mesh_vertices_2d_exterior']  # [T, 7, 2]
    mesh_vis_ext = data['mesh_vertices_vis_exterior']  # [T, 7]

    mesh_wrist = data['mesh_vertices_2d_wrist_fixed']  # [T, 7, 2]
    mesh_vis_wrist = data['mesh_vertices_vis_wrist_fixed']  # [T, 7]

    gripper_states = data['actions'][:, 6]

    # Take first 32 frames
    num_frames = min(32, len(images_ext))
    episode_frames = []

    for t in range(num_frames):
        # Exterior view
        frame_ext = images_ext[t].copy()

        # Draw mesh vertices - color by index
        for i in range(7):
            if mesh_vis_ext[t, i]:
                pt = tuple(mesh_ext[t, i].astype(int))
                if i == 0:
                    # Palm base - white
                    cv2.circle(frame_ext, pt, 4, (255, 255, 255), 2)
                elif i in [1, 3]:
                    # Finger tips - cyan
                    cv2.circle(frame_ext, pt, 5, (255, 255, 0), 2)
                elif i in [2, 4]:
                    # Finger joints - yellow
                    cv2.circle(frame_ext, pt, 3, (0, 255, 255), -1)
                else:
                    # Palm interior - green
                    cv2.circle(frame_ext, pt, 2, (0, 255, 0), -1)

        # Wrist view
        frame_wrist = images_wrist[t].copy()

        # Draw mesh vertices (hardcoded 2D - always visible, no check needed)
        for i in range(7):
            pt = tuple(mesh_wrist[t, i].astype(int))
            # Bounds check (should always pass for hardcoded coordinates)
            if 0 <= pt[0] < 320 and 0 <= pt[1] < 180:
                if i == 0:
                    # Palm base - white circle
                    cv2.circle(frame_wrist, pt, 5, (255, 255, 255), 2)
                elif i in [1, 3]:
                    # Finger tips - magenta circles (LARGER & THICKER)
                    cv2.circle(frame_wrist, pt, 7, (255, 0, 255), 3)
                elif i in [2, 4]:
                    # Finger joints - orange filled
                    cv2.circle(frame_wrist, pt, 4, (0, 165, 255), -1)
                else:
                    # Palm interior - green filled
                    cv2.circle(frame_wrist, pt, 3, (0, 255, 0), -1)

        # Combine side by side
        combined = np.concatenate([frame_ext, frame_wrist], axis=1)

        # Add text overlay
        cv2.putText(combined, f"Episode {ep_idx:03d} | Frame {t}/{num_frames} | Gripper: {gripper_states[t]:.2f}",
                   (10, 20), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 255), 2)
        cv2.putText(combined, "Exterior (3D proj)", (10, 40),
                   cv2.FONT_HERSHEY_SIMPLEX, 0.4, (255, 255, 255), 1)
        cv2.putText(combined, "Wrist (Fixed 2D)", (330, 40),
                   cv2.FONT_HERSHEY_SIMPLEX, 0.4, (255, 255, 255), 1)

        episode_frames.append(combined)

    # Save individual episode video
    output_path = output_dir / f'episode_{ep_idx:03d}_mesh_overlay.mp4'
    media.write_video(str(output_path), episode_frames, fps=10)

    data.close()

print(f"\n✓ Saved {len(samples_to_viz)} individual videos to {output_dir}")
print(f"\nEach video shows:")
print(f"  - Side-by-side: Exterior (left) | Wrist (right)")
print(f"  - Mesh vertices overlaid on actual gripper")
print(f"  - Gripper state value per frame")
print(f"\nCheck if:")
print(f"  1. Exterior: Cyan circles align with actual finger tips")
print(f"  2. Wrist: Magenta circles align with actual gripper position")
print(f"  3. Mesh follows gripper motion as gripper state changes")