reconstructions/scripts/legacy/create_multiview_data.py

# Standard Library Imports
import os
import sys
import copy
import time
import argparse
import datetime
from pathlib import Path
from collections import defaultdict, namedtuple
import struct

# Add local modules to path
sys.path.append("/home/sebastian.cavada/scsv/mast3r_complete")
sys.path.append("/home/sebastian.cavada/scsv/mast3r_complete/dust3r")
sys.path.append("/home/sebastian.cavada/scsv/mast3r_complete/_evaluation")

# Third-Party Imports
import numpy as np
import torch
import torch.nn.functional as F
import torch.backends.cudnn as cudnn
# import matplotlib.pyplot as plt
# import plotly.graph_objects as go
from PIL import Image
from tqdm import tqdm

# Local Application/Library Imports
import mast3r.utils.path_to_dust3r  # noqa
from mast3r.model import AsymmetricMASt3R
from dust3r.datasets import get_data_loader  # noqa
from dust3r.model import AsymmetricCroCo3DStereo
from dust3r.utils.geometry import geotrf, inv, normalize_pointcloud
import dust3r.datasets
import croco.utils.misc as misc  # noqa

from dust3r.inference import inference
from dust3r.image_pairs import make_pairs
from dust3r.utils.image import load_images, rgb
from dust3r.utils.device import to_numpy
from dust3r.viz import add_scene_cam, CAM_COLORS, OPENGL, pts3d_to_trimesh, cat_meshes
from dust3r.cloud_opt import global_aligner, GlobalAlignerMode

# Import utility functions
from utils.general import get_batch_colmap, plot_3d_points_with_frustums, voxel_downsample_with_colors # noqa

# creating a new dataloader
from dust3r.datasets import CustomCOLMAP

# Define infinity constant
inf = float('inf')

# Set random seed for reproducibility
seed = 777 + misc.get_rank()
torch.manual_seed(seed)
np.random.seed(seed)
cudnn.benchmark = False

# Define Point3D namedtuple for COLMAP compatibility
Point3D = namedtuple(
    "Point3D", ["id", "xyz", "rgb", "error", "image_ids", "point2D_idxs"]
)

# Default configuration
DEFAULT_CONFIG = {
    'device': 'cuda',
    'batch_size': 4,
    'schedule': 'cosine',
    'lr': 0.01,
    'niter': 300,
    'min_conf_thr': 0
}

# Model configurations
MODEL_CONFIGS = {
    'dust3r': {
        'path': '/home/sebastian.cavada/scsv/thesis/mast3r_complete/checkpoints/dust3r_512dpt/DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth',
        'architecture': "AsymmetricCroCo3DStereo(pos_embed='RoPE100', patch_embed_cls='ManyAR_PatchEmbed', img_size=(512, 512), head_type='dpt', output_mode='pts3d', depth_mode=('exp', -inf, inf), conf_mode=('exp', 1, inf), enc_embed_dim=1024, enc_depth=24, enc_num_heads=16, dec_embed_dim=768, dec_depth=12, dec_num_heads=12)"
    },
    'mast3r': {
        'path': '/home/sebastian.cavada/Documents/scsv/thesis/thesis_2025/mast3r_complete/checkpoints/dust3r_512dpt/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric.pth',
        'architecture': "AsymmetricMASt3R(pos_embed='RoPE100', patch_embed_cls='ManyAR_PatchEmbed', img_size=(512, 512), head_type='catmlp+dpt', output_mode='pts3d+desc24', depth_mode=('exp', -inf, inf), conf_mode=('exp', 1, inf), enc_embed_dim=1024, enc_depth=24, enc_num_heads=16, dec_embed_dim=768, dec_depth=12, dec_num_heads=12, two_confs=True, desc_conf_mode=('exp', 0, inf), use_intrinsics=False, use_extrinsics=False)"
    }
}

def read_next_bytes(fid, num_bytes, format_char_sequence, endian_character="<"):
    """Read and unpack the next bytes from a binary file.
    :param fid:
    :param num_bytes: Sum of combination of {2, 4, 8}, e.g. 2, 6, 16, 30, etc.
    :param format_char_sequence: List of {c, e, f, d, h, H, i, I, l, L, q, Q}.
    :param endian_character: Any of {@, =, <, >, !}
    :return: Tuple of read and unpacked values.
    """
    data = fid.read(num_bytes)
    return struct.unpack(endian_character + format_char_sequence, data)

def read_points3D_binary(path_to_model_file):
    """
    Read Points3D from COLMAP binary file.
    
    Args:
        path_to_model_file: Path to COLMAP points3D.bin file
        
    Returns:
        points3D: Dictionary of Point3D objects
    """
    points3D = {}
    with open(path_to_model_file, "rb") as fid:
        num_points = read_next_bytes(fid, 8, "Q")[0]
        for _ in range(num_points):
            binary_point_line_properties = read_next_bytes(
                fid, num_bytes=43, format_char_sequence="QdddBBBd"
            )
            point3D_id = binary_point_line_properties[0]
            xyz = np.array(binary_point_line_properties[1:4])
            rgb = np.array(binary_point_line_properties[4:7])
            error = np.array(binary_point_line_properties[7])
            track_length = read_next_bytes(
                fid, num_bytes=8, format_char_sequence="Q"
            )[0]
            track_elems = read_next_bytes(
                fid,
                num_bytes=8 * track_length,
                format_char_sequence="ii" * track_length,
            )
            image_ids = np.array(tuple(map(int, track_elems[0::2])))
            point2D_idxs = np.array(tuple(map(int, track_elems[1::2])))
            points3D[point3D_id] = Point3D(
                id=point3D_id,
                xyz=xyz,
                rgb=rgb,
                error=error,
                image_ids=image_ids,
                point2D_idxs=point2D_idxs,
            )
    return points3D

def write_next_bytes(fid, data, format_char_sequence, endian_character="<"):
    """Pack and write to a binary file.
    :param fid:
    :param data: data to send, if multiple elements are sent at the same time,
    they should be encapsuled either in a list or a tuple
    :param format_char_sequence: List of {c, e, f, d, h, H, i, I, l, L, q, Q}.
    should be the same length as the data list or tuple
    :param endian_character: Any of {@, =, <, >, !}
    """
    if isinstance(data, (list, tuple)):
        bytes = struct.pack(endian_character + format_char_sequence, *data)
    else:
        bytes = struct.pack(endian_character + format_char_sequence, data)
    fid.write(bytes)

def write_points3D_binary(points3D, path_to_model_file):
    """
    Write Points3D to COLMAP binary file.
    
    Args:
        points3D: Dictionary of Point3D objects
        path_to_model_file: Path to write COLMAP points3D.bin file
    """
    with open(path_to_model_file, "wb") as fid:
        write_next_bytes(fid, len(points3D), "Q")
        for _, pt in points3D.items():
            write_next_bytes(fid, pt.id, "Q")
            write_next_bytes(fid, pt.xyz.tolist(), "ddd")
            write_next_bytes(fid, pt.rgb.tolist(), "BBB")
            write_next_bytes(fid, pt.error, "d")
            track_length = pt.image_ids.shape[0]
            write_next_bytes(fid, track_length, "Q")
            for image_id, point2D_id in zip(pt.image_ids, pt.point2D_idxs):
                write_next_bytes(fid, [image_id, point2D_id], "ii")

def load_model(model_name, device):
    """
    Load the specified model.

    Args:
        model_name: Name of the model to load ('dust3r', 'mast3r')
        device: Device to load the model on

    Returns:
        loaded_model: The loaded model
    """
    if model_name not in MODEL_CONFIGS:
        raise ValueError(f"Model {model_name} not supported. Choose from: {list(MODEL_CONFIGS.keys())}")

    config = MODEL_CONFIGS[model_name]
    model_path = config['path']
    model_architecture = config['architecture']

    # Create model instance based on architecture string
    model = eval(model_architecture)
    model.to(device)

    print(f'Loading pretrained model: {model_path}')
    checkpoint = torch.load(model_path, map_location=device)
    model.load_state_dict(checkpoint['model'], strict=False)
    model.eval()

    return model

def get_transformation_between_cameras(target_poses, source_poses):
    """
    Calculate transformation between source and target camera poses.
    
    Args:
        target_poses: Target camera poses (ground truth)
        source_poses: Source camera poses (predicted)
        
    Returns:
        scale: Scale factor
        rotation_matrix: Rotation matrix
        translation_shift: Translation vector
    """
    source_translation_1 = source_poses[0][:3, 3]
    source_translation_2 = source_poses[1][:3, 3]

    source_rotation_1 = source_poses[0][:3, :3]
    source_rotation_2 = source_poses[1][:3, :3]

    target_translation_1 = target_poses[0][:3, 3]
    target_translation_2 = target_poses[1][:3, 3]

    target_rotation_1 = target_poses[0][:3, :3]
    target_rotation_2 = target_poses[1][:3, :3]

    # 1. Calculate scale: ratio of distances between corresponding positions (target vs source)
    source_distance = np.linalg.norm(source_translation_2 - source_translation_1)
    target_distance = np.linalg.norm(target_translation_2 - target_translation_1)
    
    scale = target_distance / source_distance

    # 2. Compute rotation matrix to align source with target
    rotation_matrix = source_rotation_1 @ np.linalg.inv(target_rotation_1)

    # 3. Calculate translation shift
    # since the first camera is always in the origin:
    translation_shift = target_translation_1 - source_translation_1

    return scale, rotation_matrix, translation_shift

def process_scene(scene_config, dataset, model, device):
    """
    Process a scene with the given configuration.
    
    Args:
        scene_config: Configuration for the scene
        dataset: Dataset object
        model: Model to use for inference
        device: Device to run inference on
        
    Returns:
        points_cat: List of transformed point clouds
        colors_cat: List of point colors
        cameras_cat: List of camera poses
    """
    # Extract scene configuration
    selected_indices = scene_config.get('selected_indices', [])
    if not selected_indices and scene_config.get('auto_select', False):
        # Generate indices automatically based on configuration
        num_images = scene_config.get('num_images', 120)
        distance = scene_config.get('distance', 10)
        offset = scene_config.get('offset', 0)
        step = scene_config.get('step', 3)
        
        selected_indices = [(i + j) + offset for i in range(0, num_images, distance) 
                           for j in (0, step) if i + j < num_images]
    
    # Make sure we have an even number of indices for pair processing
    if len(selected_indices) % 2 != 0:
        selected_indices = selected_indices[:-1]
        
    print(f"Processing scene with indices: {selected_indices}")
    
    # Get batch of data for selected indices
    batch = get_batch_colmap(dataset, 
                            scene_config.get('max_batch_size', 50), 
                            scene_id=scene_config.get('scene_id', 0), 
                            selected_indices=selected_indices)
    
    print(f"Batch size: {len(batch)}")
    
    predictions = []
    camera_poses_gt_from_pred = []
    
    # Process each pair of images
    for i in tqdm(range(0, len(selected_indices), 2), desc="Processing pairs"):
        print(f"Processing pair {i//2+1}/{len(selected_indices)//2}")
        pair_batch = batch[i:i+2]
        
        # Load images and make pairs
        images = load_images([pair_batch[0]['path'][0], pair_batch[1]['path'][0]], size=512)
        pairs = make_pairs(images, scene_graph='complete', prefilter=None, symmetrize=True)    
        
        # Run inference
        output = inference(pairs, model, device, batch_size=scene_config.get('batch_size', 32))
        
        # Global alignment
        scene = global_aligner(output, device=device, mode=GlobalAlignerMode.PointCloudOptimizer)
        scene.compute_global_alignment(
            init="mst", 
            niter=scene_config.get('niter', 100), 
            schedule=scene_config.get('schedule', 'cosine'), 
            lr=scene_config.get('lr', 0.001)
        )
        
        # Post-processing
        scene.mask_sky()
        scene.clean_pointcloud()
        predictions.append(scene)
        
        # Get ground truth camera poses
        gt_poses_0 = pair_batch[0]['camera_pose']
        gt_poses_1 = pair_batch[1]['camera_pose']
        gt_poses = torch.cat([gt_poses_0, gt_poses_1], dim=0).cpu()
        camera_poses_gt_from_pred.append(gt_poses)
    
    # Combine results
    points_cat = []
    colors_cat = []
    cameras_cat = []
    
    for i, scene in enumerate(predictions):
        rgbimg = scene.imgs
        pts3d = to_numpy(scene.get_pts3d())
        mask = to_numpy(scene.get_masks())
        im_poses = scene.get_im_poses()
        
        # Concatenate points and colors
        points = np.concatenate([p[m] for p, m in zip(pts3d, mask)])
        colors = np.concatenate([p[m] for p, m in zip(rgbimg, mask)]) * 255
        
        # Transform to align with ground truth
        scale, rotation, translation = get_transformation_between_cameras(
            camera_poses_gt_from_pred[i].detach().cpu(), 
            im_poses.detach().cpu()
        )
        
        # Apply transformation to points
        points_scaled = points * scale
        points_translated = points_scaled @ rotation.numpy() + translation.numpy()
        
        points_cat.append(points_translated)
        colors_cat.append(colors)
        cameras_cat.append(camera_poses_gt_from_pred[i])
    
    return points_cat, colors_cat, cameras_cat

def main():
    """
    Main function to run the multi-scene reconstruction pipeline.
    """
    # Define scene configurations
    scene_configs = [
        # {
        #     'name': 'campus_hydro',
        #     'selected_indices': [0, 5, 10, 15, 20, 25, 30, 35]        
        # },
        {
            'name': 'campus_core',
            'selected_indices': [0, 5, 10, 15, 20, 25, 30, 35],
            'voxel_size': 0.05

        }
        # {
        #     'name': 'campus_hydro_secondary',
        #     'scene_id': 0,
        #     'selected_indices': [0, 5, 10, 15, 20, 25, 30, 35],
        #     'batch_size': 32,
        #     'schedule': 'cosine',
        #     'lr': 0.001,
        #     'niter': 100,
        #     'min_conf_thr': 0.5,
        #     'voxel_size': 0.05
        # }
    ]
    
    # Setup paths
    base_path = "/l/users/sebastian.cavada/MBZUAI-Campus/global_OG"
    data_path = f"{base_path}/_data_global"
    output_path = f"{base_path}/_data_global_denser"
    
    # Parse arguments (using default values for now)
    args = argparse.Namespace(
        model='mast3r',
        device='cuda',
        batch_size=32
    )        
    
    # Load model
    model = load_model(args.model, args.device)
    
    # Process each scene
    all_points = []
    all_colors = []
    
    for scene_config in scene_configs:
        print(f"\nLoading dataset for scene: {scene_config['name']}")
        # Load dataset
        dataset = CustomCOLMAP(
            size=200,
            split='train',
            images_path=f"{data_path}/images/",
            sfm_path=f"{data_path}/{scene_config['name']}/0",
            resolution=(512, 384),
            seed=777,
        )


        print(f"\nProcessing scene: {scene_config['name']}")
        
        # Process scene
        points_cat, colors_cat, cameras_cat = process_scene(
            scene_config, dataset, model, args.device
        )
        
        # Combine points and colors
        if points_cat:
            scene_points = np.concatenate(points_cat, axis=0)
            scene_colors = np.concatenate(colors_cat, axis=0)
            
            # Downsample if needed
            if scene_config.get('voxel_size', 0) > 0:
                scene_points, scene_colors, _ = voxel_downsample_with_colors(
                    scene_points, scene_colors, voxel_size=scene_config.get('voxel_size', 0.05)
                )
            
            all_points.append(scene_points)
            all_colors.append(scene_colors)
            
            print(f"Added {len(scene_points)} points from scene {scene_config['name']}")
    
    # Combine all scenes
    if all_points:
        points_total = np.concatenate(all_points, axis=0)
        colors_total = np.concatenate(all_colors, axis=0)
        
        print(f"Total points before final downsampling: {len(points_total)}")
        
        # Final downsampling
        total_points_downsampled, total_colors_downsampled, _ = voxel_downsample_with_colors(
            points_total, colors_total, voxel_size=0.05
        )
        
        print(f"Total points after final downsampling: {len(total_points_downsampled)}")
        
        # Read existing COLMAP points
        output_dir = f"{output_path}/campus_hydro"
        os.makedirs(output_dir, exist_ok=True)
        
        try:
            points_colmap = read_points3D_binary(f"{data_path}/campus_hydro/0/points3D.bin")
            print(f"Loaded {len(points_colmap)} existing points from COLMAP")
        except FileNotFoundError:
            points_colmap = {}
            print("No existing COLMAP points found, starting with empty set")
        
        # Find highest ID
        highest_id = max(points_colmap.keys()) if points_colmap else 0
        
        # Add new points
        for xyz, rgb in zip(total_points_downsampled, total_colors_downsampled):
            xyz_np = np.array(xyz)
            rgb_int = rgb.astype(int)
            
            highest_id += 1
            points_colmap[highest_id] = Point3D(
                id=highest_id,
                xyz=xyz_np,
                rgb=rgb_int,
                error=0,
                image_ids=np.array([]),
                point2D_idxs=np.array([])
            )
        
        print(f"Final point count: {len(points_colmap)}")
        
        # Write points to file
        write_points3D_binary(points_colmap, f"{output_dir}/points3D.bin")
        print(f"Points written to {output_dir}/points3D.bin")
    else:
        print("No points were processed. Check scene configurations.")

if __name__ == "__main__":
    main()