2025 update

README.md

@@ -1,20 +1,77 @@
 ---
 license: apache-2.0
 ---
-# HoHo Tools
+# HoHo2025 Tools
 
-Tools and utilities for the [S23DR competition](https://huggingface.co/spaces/usm3d/S23DR) and [HoHo Dataset](https://huggingface.co/datasets/usm3d/usm-training-data)
+Tools and utilities for the [S23DR-2025 competition](https://huggingface.co/spaces/usm3d/S23DR2025) and [HoHo25k Dataset](https://huggingface.co/datasets/usm3d/hoho25k)
 
 ## Installation 
-```bash
-# pip install over ssh
-pip install git+ssh://git@hf.co/usm3d/tools2025.git 
 
-# pip install over http
+### pip install over http
+```bash
 pip install git+http://hf.co/usm3d/tools2025.git    
+```
 
-# editable
+or editable 
+```bash
 git clone http://hf.co/usm3d/tools2025 
-cd tools
+cd tools2025
 pip install -e .
+```
+
+### Usage example
+
+```python
+from datasets import load_dataset
+from hoho2025.vis import plot_all_modalities
+from hoho2025.viz3d import *
+
+def read_colmap_rec(colmap_data):
+    import pycolmap
+    import tempfile,zipfile
+    import io
+    with tempfile.TemporaryDirectory() as tmpdir:
+        with zipfile.ZipFile(io.BytesIO(colmap_data), "r") as zf:
+            zf.extractall(tmpdir)  # unpacks cameras.txt, images.txt, etc. to tmpdir
+        # Now parse with pycolmap
+        rec = pycolmap.Reconstruction(tmpdir)
+        return rec
+
+ds = load_dataset("usm3d/hoho25k", streaming=True, trust_remote_code=True)
+for a in ds['train']:
+    break
+
+fig, ax = plot_all_modalities(a)
+
+## Now 3d
+
+fig3d = init_figure()
+plot_reconstruction(fig3d, read_colmap_rec(a['colmap_binary']))
+plot_wireframe(fig3d, a['wf_vertices'], a['wf_edges'], a['wf_classifications'])
+plot_bpo_cameras_from_entry(fig3d, a)
+fig3d
+```
+
+## Example wireframe estimation 
+
+Look in [hoho2025/example_solution.py](hoho2025/example_solution.py)
+
+```python
+from hoho2025.example_solutions import predict_wireframe
+pred_vertices, pred_connections = predict_wireframe(a)
+
+fig3d = init_figure()
+plot_reconstruction(fig3d, read_colmap_rec(a['colmap_binary']))
+plot_wireframe(fig3d, pred_vertices, pred_connections, color='rgb(0, 0, 255)')
+fig3d
+```
+
+
+And to get the metric
+
+```python
+from hoho2025.metric_helper import hss
+
+score = hss(pred_vertices, pred_connections, a['wf_vertices'], a['wf_edges'], vert_thresh=0.5, edge_thresh=0.5)
+print (score)
 ```

hoho/wed.py

@@ -1,107 +0,0 @@
-from scipy.spatial.distance import cdist
-from scipy.optimize import linear_sum_assignment
-import numpy as np 
-
-
-def preregister_mean_std(verts_to_transform, target_verts, single_scale=True):
-    mu_target = target_verts.mean(axis=0)
-    mu_in = verts_to_transform.mean(axis=0)
-    std_target = np.std(target_verts, axis=0)
-    std_in = np.std(verts_to_transform, axis=0)
-    
-    if np.any(std_in == 0):
-        std_in[std_in == 0] = 1
-    if np.any(std_target == 0):
-        std_target[std_target == 0] = 1
-    if np.any(np.isnan(std_in)):
-        std_in[np.isnan(std_in)] = 1
-    if np.any(np.isnan(std_target)):
-        std_target[np.isnan(std_target)] = 1
-        
-    if single_scale:
-        std_target = np.linalg.norm(std_target)
-        std_in = np.linalg.norm(std_in)
-    
-    transformed_verts = (verts_to_transform - mu_in) / std_in
-    transformed_verts = transformed_verts * std_target + mu_target
-    
-    return transformed_verts
-
-
-def update_cv(cv, gt_vertices):
-    if cv < 0:
-        diameter = cdist(gt_vertices, gt_vertices).max()
-        # Cost of adding or deleting a vertex is set to -cv times the diameter of the ground truth wireframe
-        cv = -cv * diameter
-    return cv
-
-def compute_WED(pd_vertices, pd_edges, gt_vertices, gt_edges, cv_ins=-1/2, cv_del=-1/4, ce=1.0, normalized=True, preregister=True, single_scale=True):
-    '''The function computes the Wireframe Edge Distance (WED) between two graphs.
-    pd_vertices: list of predicted vertices
-    pd_edges: list of predicted edges
-    gt_vertices: list of ground truth vertices
-    gt_edges: list of ground truth edges
-    cv_ins: vertex insertion cost: if positive, the cost in centimeters of inserting vertex, if negative, multiplies diameter to compute cost (default is -1/2)
-    cv_del: vertex deletion cost: if positive, the cost in centimeters of deleting a vertex, if negative, multiplies diameter to compute cost (default is -1/4)
-    ce: edge cost (multiplier of the edge length for edge deletion and insertion, default is 1.0)
-    normalized: if True, the WED is normalized by the total length of the ground truth edges
-    preregister: if True, the predicted vertices have their mean and scale matched to the ground truth vertices
-    '''
-    
-    pd_vertices = np.array(pd_vertices)
-    gt_vertices = np.array(gt_vertices)
-    pd_edges = np.array(pd_edges)
-    gt_edges = np.array(gt_edges)        
-    
-        
-    cv_del = update_cv(cv_del, gt_vertices)
-    cv_ins = update_cv(cv_ins, gt_vertices)
-    
-    # Step 0: Prenormalize / preregister
-    if preregister:
-        pd_vertices = preregister_mean_std(pd_vertices, gt_vertices, single_scale=single_scale)
-    
-
-    # Step 1: Bipartite Matching
-    distances = cdist(pd_vertices, gt_vertices, metric='euclidean')
-    row_ind, col_ind = linear_sum_assignment(distances)
-        
-    
-    # Step 2: Vertex Translation
-    translation_costs = np.sum(distances[row_ind, col_ind])
-    
-    # Step 3: Vertex Deletion
-    unmatched_pd_indices = set(range(len(pd_vertices))) - set(row_ind)
-    deletion_costs = cv_del * len(unmatched_pd_indices)  
-    
-    # Step 4: Vertex Insertion
-    unmatched_gt_indices = set(range(len(gt_vertices))) - set(col_ind)
-    insertion_costs = cv_ins * len(unmatched_gt_indices)  
-    
-    # Step 5: Edge Deletion and Insertion
-    updated_pd_edges = [(col_ind[np.where(row_ind == edge[0])[0][0]], col_ind[np.where(row_ind == edge[1])[0][0]]) for edge in pd_edges if len(edge)==2 and edge[0] in row_ind and edge[1] in row_ind]
-    pd_edges_set = set(map(tuple, [set(edge) for edge in updated_pd_edges]))
-    gt_edges_set = set(map(tuple, [set(edge) for edge in gt_edges]))
-
-    
-    # Delete edges not in ground truth
-    edges_to_delete = pd_edges_set - gt_edges_set
-    
-    vert_tf = [np.where(col_ind == v)[0][0] if v in col_ind else 0 for v in range(len(gt_vertices))]
-    deletion_edge_costs = ce * sum(np.linalg.norm(pd_vertices[vert_tf[edge[0]]] - pd_vertices[vert_tf[edge[1]]]) for edge in edges_to_delete if len(edge) == 2)
-
-    
-    # Insert missing edges from ground truth
-    edges_to_insert = gt_edges_set - pd_edges_set
-    insertion_edge_costs = ce * sum(np.linalg.norm(gt_vertices[edge[0]] - gt_vertices[edge[1]]) for edge in edges_to_insert if len(edge) == 2) 
-    
-    # Step 6: Calculation of WED
-    WED = translation_costs + deletion_costs + insertion_costs + deletion_edge_costs + insertion_edge_costs
-
-    
-    if normalized:
-        total_length_of_gt_edges = np.linalg.norm((gt_vertices[gt_edges[:, 0]] - gt_vertices[gt_edges[:, 1]]), axis=1).sum()
-        WED = WED / total_length_of_gt_edges
-        
-    # print ("Total length", total_length_of_gt_edges)
-    return WED

{hoho → hoho2025}/__init__.py

@@ -1,7 +1,5 @@
 from .hoho import *
 from . import vis
-from . import read_write_colmap
-from .wed import compute_WED
 
 import importlib
 import sys

{hoho → hoho2025}/color_mappings.py

@@ -1,5 +1,3 @@
-import numpy as np 
-
 gestalt_color_mapping = {
     "unclassified": (215, 62, 138),
     "apex": (235, 88, 48),
@@ -185,22 +183,27 @@ ade20k_color_mapping = {
 }
 
 
-# edge_colors = np.asarray([(214, 251, 248),
-#                           (13, 94, 47),
-#                           (54, 243, 63),
-#                           (187, 123, 236),
-#                           (162, 162, 32),
-#                           (169, 255, 219),
-#                           (8, 89, 52),
-#                           (85, 27, 65),
-#                           (0, 0, 0)]
-
+EDGE_CLASSES = {'cornice_return': 0,
+                'cornice_strip': 1,
+                'eave': 2,
+                'flashing': 3,
+                'hip': 4,
+                'rake': 5,
+                'ridge': 6,
+                'step_flashing': 7,
+                'transition_line': 8,
+                'valley': 9}
+EDGE_CLASSES_BY_ID = {v: k for k, v in EDGE_CLASSES.items()}
 
-# edge_colors = np.array([[ 54, 243,  63], 
-#                         [214, 251, 248], 
-#                         [169, 255, 219], 
-#                         [ 13,  94,  47], 
-#                         [162, 162,  32], 
-#                         [187, 123, 236], 
-#                         [ 85,  27,  65], 
-#                         [  0,  0,    0]])
+edge_color_mapping = {
+    'cornice_return': (215, 62, 138),
+    'cornice_strip': (235, 88, 48),
+    'eave':  (54, 243, 63),
+    "flashing": (162, 162, 32),
+    'hip': (8, 89, 52),
+    'rake': (13, 94, 47),
+    'ridge': (214, 251, 248),
+    "step_flashing": (169, 255, 219),
+    'transition_line': (200,0,50),
+    'valley': (85, 27, 65),
+}

hoho2025/example_solutions.py

@@ -0,0 +1,701 @@
+# Description: This file contains the handcrafted solution for the task of wireframe reconstruction 
+import io
+import tempfile
+import zipfile
+from collections import defaultdict
+from typing import Tuple, List
+import cv2
+import numpy as np
+import pycolmap
+from PIL import Image as PImage
+from scipy.spatial.distance import cdist
+
+from hoho2025.color_mappings import ade20k_color_mapping, gestalt_color_mapping
+
+
+def empty_solution():
+    '''Return a minimal valid solution, i.e. 2 vertices and 1 edge.'''
+    return np.zeros((2,3)), [(0, 1)]
+    
+    
+def read_colmap_rec(colmap_data):
+    with tempfile.TemporaryDirectory() as tmpdir:
+        with zipfile.ZipFile(io.BytesIO(colmap_data), "r") as zf:
+            zf.extractall(tmpdir)  # unpacks cameras.txt, images.txt, etc. to tmpdir
+        # Now parse with pycolmap
+        rec = pycolmap.Reconstruction(tmpdir)
+        return rec
+
+def convert_entry_to_human_readable(entry):
+    out = {}
+    for k, v in entry.items():
+        if 'colmap' in k:
+            out[k] = read_colmap_rec(v)
+        elif k in ['wf_vertices', 'wf_edges', 'K', 'R', 't', 'depth']:
+            out[k] = np.array(v)
+        else:
+            out[k]=v
+    out['__key__'] = entry['order_id']
+    return out
+
+
+def get_house_mask(ade20k_seg):
+    """
+    Get a mask of the house in the ADE20K segmentation map.
+    """
+    house_classes_ade20k = [
+        'wall',
+        'house',
+        'building;edifice',
+        'door;double;door',
+        'windowpane;window',
+    ]
+    np_seg = np.array(ade20k_seg)
+    full_mask = np.zeros(np_seg.shape[:2], dtype=np.uint8)
+    for c in house_classes_ade20k:
+        color = np.array(ade20k_color_mapping[c])
+        mask = cv2.inRange(np_seg, color-0.5, color+0.5)
+        full_mask = np.logical_or(full_mask, mask)
+    return full_mask
+
+
+def point_to_segment_dist(pt, seg_p1, seg_p2):
+    """
+    Computes the Euclidean distance from pt to the line segment p1->p2.
+    pt, seg_p1, seg_p2: (x, y) as np.ndarray
+    """
+    # If both endpoints are the same, just return distance to one of them
+    if np.allclose(seg_p1, seg_p2):
+        return np.linalg.norm(pt - seg_p1)
+    seg_vec = seg_p2 - seg_p1
+    pt_vec = pt - seg_p1
+    seg_len2 = seg_vec.dot(seg_vec)
+    t = max(0, min(1, pt_vec.dot(seg_vec)/seg_len2))
+    proj = seg_p1 + t*seg_vec
+    return np.linalg.norm(pt - proj)
+
+
+def get_vertices_and_edges_from_segmentation(gest_seg_np, edge_th=25.0):
+    """
+    Identify apex and eave-end vertices, then detect lines for eave/ridge/rake/valley.
+    For each connected component, we do a line fit with cv2.fitLine, then measure
+    segment endpoints more robustly. We then associate apex points that are within
+    'edge_th' of the line segment. We record those apex–apex connections for edges
+    if at least 2 apexes lie near the same component line.
+    """
+    #--------------------------------------------------------------------------------
+    # Step A: Collect apex and eave_end vertices
+    #--------------------------------------------------------------------------------
+    if not isinstance(gest_seg_np, np.ndarray):
+        gest_seg_np = np.array(gest_seg_np)
+    vertices = []
+    # Apex
+    apex_color = np.array(gestalt_color_mapping['apex'])
+    apex_mask = cv2.inRange(gest_seg_np, apex_color-0.5, apex_color+0.5)
+    if apex_mask.sum() > 0:
+        output = cv2.connectedComponentsWithStats(apex_mask, 8, cv2.CV_32S)
+        (numLabels, labels, stats, centroids) = output
+        stats, centroids = stats[1:], centroids[1:]  # skip background
+        for i in range(numLabels-1):
+            vert = {"xy": centroids[i], "type": "apex"}
+            vertices.append(vert)
+
+    # Eave end
+    eave_end_color = np.array(gestalt_color_mapping['eave_end_point'])
+    eave_end_mask = cv2.inRange(gest_seg_np, eave_end_color-0.5, eave_end_color+0.5)
+    if eave_end_mask.sum() > 0:
+        output = cv2.connectedComponentsWithStats(eave_end_mask, 8, cv2.CV_32S)
+        (numLabels, labels, stats, centroids) = output
+        stats, centroids = stats[1:], centroids[1:]
+        for i in range(numLabels-1):
+            vert = {"xy": centroids[i], "type": "eave_end_point"}
+            vertices.append(vert)
+
+    # Consolidate apex points as array:
+    apex_pts = []
+    apex_idx_map = []  # keep track of index in 'vertices'
+    for idx, v in enumerate(vertices):
+        apex_pts.append(v['xy'])
+        apex_idx_map.append(idx)
+    apex_pts = np.array(apex_pts)
+
+    connections = []
+    edge_classes = ['eave', 'ridge', 'rake', 'valley']
+    for edge_class in edge_classes:
+        edge_color = np.array(gestalt_color_mapping[edge_class])
+        mask_raw = cv2.inRange(gest_seg_np, edge_color-0.5, edge_color+0.5)
+        # Possibly do morphological open/close to avoid merges or small holes
+        kernel = np.ones((5, 5), np.uint8)  # smaller kernel to reduce over-merge
+        mask = cv2.morphologyEx(mask_raw, cv2.MORPH_CLOSE, kernel)
+        if mask.sum() == 0:
+            continue
+
+        # Connected components
+        output = cv2.connectedComponentsWithStats(mask, 8, cv2.CV_32S)
+        (numLabels, labels, stats, centroids) = output
+        # skip the background
+        stats, centroids = stats[1:], centroids[1:]
+        label_indices = range(1, numLabels)
+
+        # For each connected component, do a line fit
+        for lbl in label_indices:
+            ys, xs = np.where(labels == lbl)
+            if len(xs) < 2:
+                continue
+            # Fit a line using cv2.fitLine
+            pts_for_fit = np.column_stack([xs, ys]).astype(np.float32)
+            # (vx, vy, x0, y0) = direction + a point on the line
+            line_params = cv2.fitLine(pts_for_fit, distType=cv2.DIST_L2, 
+                                      param=0, reps=0.01, aeps=0.01)
+            vx, vy, x0, y0 = line_params.ravel()
+            # We'll approximate endpoints by projecting (xs, ys) onto the line,
+            # then taking min and max in the 1D param along the line.
+
+            # param along the line = ( (x - x0)*vx + (y - y0)*vy )
+            proj = ( (xs - x0)*vx + (ys - y0)*vy )
+            proj_min, proj_max = proj.min(), proj.max()
+            p1 = np.array([x0 + proj_min*vx, y0 + proj_min*vy])
+            p2 = np.array([x0 + proj_max*vx, y0 + proj_max*vy])
+
+            #--------------------------------------------------------------------------------
+            # Step C: If apex points are within 'edge_th' of segment, they are connected
+            #--------------------------------------------------------------------------------
+            if len(apex_pts) < 2:
+                continue
+
+            # Distance from each apex to the line segment
+            dists = np.array([
+                point_to_segment_dist(apex_pts[i], p1, p2)
+                for i in range(len(apex_pts))
+            ])
+
+            # Indices of apex points that are near
+            near_mask = (dists <= edge_th)
+            near_indices = np.where(near_mask)[0]
+            if len(near_indices) < 2:
+                continue
+
+            # Connect each pair among these near apex points
+            for i in range(len(near_indices)):
+                for j in range(i+1, len(near_indices)):
+                    a_idx = near_indices[i]
+                    b_idx = near_indices[j]
+                    # 'a_idx' and 'b_idx' are indices in apex_pts / apex_idx_map
+                    vA = apex_idx_map[a_idx]
+                    vB = apex_idx_map[b_idx]
+                    # Store the connection using sorted indexing
+                    conn = tuple(sorted((vA, vB)))
+                    connections.append(conn)
+
+    return vertices, connections
+
+
+def get_uv_depth(vertices: List[dict],
+                 depth_fitted: np.ndarray,
+                 sparse_depth: np.ndarray,
+                 search_radius: int = 10) -> Tuple[np.ndarray, np.ndarray]:
+    """
+    For each vertex, returns a 2D array of (u,v) and a matching 1D array of depths.
+    
+    We attempt to use the sparse_depth if available in a local neighborhood:
+      1. For each vertex coordinate (x, y), define a local window in sparse_depth 
+         of size (2*search_radius + 1).
+      2. Collect all valid (nonzero) values in that window.
+      3. If any exist, we take the *closest* valid pixel's depth.
+      4. Otherwise, we use depth_fitted[y, x].
+    
+    Parameters
+    ----------
+    vertices : List[dict]
+        Each dict must have "xy" at least, e.g. {"xy": (x, y), ...}
+    depth_fitted : np.ndarray
+        A 2D array (H, W), the dense (or corrected) depth for fallback.
+    sparse_depth : np.ndarray
+        A 2D array (H, W), mostly zeros except where accurate data is available.
+    search_radius : int
+        Pixel radius around the vertex in which to look for sparse depth values.
+    
+    Returns
+    -------
+    uv : np.ndarray of shape (N, 2)
+        2D float coordinates of each vertex (x, y).
+    vertex_depth : np.ndarray of shape (N,)
+        Depth value chosen for each vertex.
+    """
+    
+    # Collect each vertex's (x, y)
+    uv = np.array([vert['xy'] for vert in vertices], dtype=np.float32)
+    
+    # Convert to integer pixel coordinates (round or floor)
+    uv_int = np.round(uv).astype(np.int32)
+    H, W = depth_fitted.shape[:2]
+    
+    # Clip coordinates to stay within image bounds
+    uv_int[:, 0] = np.clip(uv_int[:, 0], 0, W - 1)
+    uv_int[:, 1] = np.clip(uv_int[:, 1], 0, H - 1)
+    
+    # Prepare output array of depths
+    vertex_depth = np.zeros(len(vertices), dtype=np.float32)
+    dense_count = 0
+    
+    for i, (x_i, y_i) in enumerate(uv_int):
+        # Local region in [x_i - search_radius, x_i + search_radius]
+        x0 = max(0, x_i - search_radius)
+        x1 = min(W, x_i + search_radius + 1)
+        y0 = max(0, y_i - search_radius)
+        y1 = min(H, y_i + search_radius + 1)
+        
+        # Crop out the local window in sparse_depth
+        region = sparse_depth[y0:y1, x0:x1]
+        
+        # Find all valid (non-zero) depths
+        valid_mask = (region > 0)
+        valid_y, valid_x = np.where(valid_mask)
+        
+        if valid_y.size > 0:
+            # Compute global coordinates for each valid pixel
+            global_x = x0 + valid_x
+            global_y = y0 + valid_y
+            
+            # Compute squared distance to center (x_i, y_i)
+            dist_sq = (global_x - x_i)**2 + (global_y - y_i)**2
+            
+            # Find the nearest valid pixel
+            min_idx = np.argmin(dist_sq)
+            nearest_depth = region[valid_y[min_idx], valid_x[min_idx]]
+            vertex_depth[i] = nearest_depth
+        else:
+            # Fallback to the dense depth
+            vertex_depth[i] = depth_fitted[y_i, x_i]
+            dense_count += 1
+    return uv, vertex_depth
+
+
+
+def project_vertices_to_3d(uv: np.ndarray, depth_vert: np.ndarray, col_img: pycolmap.Image) -> np.ndarray:
+    """
+    Projects 2D vertex coordinates with associated depths to 3D world coordinates.
+
+    Parameters
+    ----------
+    uv : np.ndarray
+        (N, 2) array of 2D vertex coordinates (u, v).
+    depth_vert : np.ndarray
+        (N,) array of depth values for each vertex.
+    col_img : pycolmap.Image
+
+    Returns
+    -------
+    vertices_3d : np.ndarray
+        (N, 3) array of vertex coordinates in 3D world space.
+    """
+    # Backproject to 3D local camera coordinates
+    xy_local = np.ones((len(uv), 3))
+    K = col_img.camera.calibration_matrix()
+    xy_local[:, 0] = (uv[:, 0] - K[0, 2]) / K[0, 0]
+    xy_local[:, 1] = (uv[:, 1] - K[1, 2]) / K[1, 1]
+    # Get the 3D vertices
+    vertices_3d_local = xy_local * depth_vert[...,None]
+    
+    # Create camera-to-world transformation matrix
+    world_to_cam = np.eye(4)
+    world_to_cam[:3] = col_img.cam_from_world.matrix()
+    cam_to_world = np.linalg.inv(world_to_cam)
+    
+    # Transform local 3D points to world coordinates
+    vertices_3d_homogeneous = cv2.convertPointsToHomogeneous(vertices_3d_local)
+    vertices_3d = cv2.transform(vertices_3d_homogeneous, cam_to_world)
+    vertices_3d = cv2.convertPointsFromHomogeneous(vertices_3d).reshape(-1, 3)
+    return vertices_3d
+
+
+def create_3d_wireframe_single_image(vertices: List[dict],
+                                     connections: List[Tuple[int, int]],
+                                     depth: PImage,
+                                     colmap_rec: pycolmap.Reconstruction,
+                                     img_id: str,
+                                     ade_seg: PImage) -> np.ndarray:
+    """
+    Processes a single image view to generate 3D vertex coordinates from existing 2D vertices/edges.
+
+    Parameters
+    ----------
+    vertices : List[dict]
+        List of 2D vertex dictionaries (e.g., {"xy": (x, y), "type": ...}).
+    connections : List[Tuple[int, int]]
+        List of 2D edge connections (indices into the vertices list).
+    depth : PIL.Image
+        Initial dense depth map as a PIL Image.
+    colmap_rec : pycolmap.Reconstruction
+        COLMAP reconstruction data.
+    img_id : str
+        Identifier for the current image within the COLMAP reconstruction.
+    ade_seg : PIL.Image
+        ADE20k segmentation map for the image.
+
+    Returns
+    -------
+    vertices_3d : np.ndarray
+        (N, 3) array of vertex coordinates in 3D world space.
+        Returns an empty array if processing fails (e.g., missing sparse depth).
+    """
+    # Check if initial vertices/connections are valid
+    if (len(vertices) < 2) or (len(connections) < 1):
+        # This case should ideally be handled before calling, but good to double check.
+        print(f'Warning: create_3d_wireframe_single_image called with insufficient vertices/connections for image {img_id}')
+        return np.empty((0, 3))
+
+    # Get fitted dense depth and sparse depth
+    depth_fitted, depth_sparse, found_sparse, col_img = get_fitted_dense_depth(
+        depth, colmap_rec, img_id, ade_seg
+    )
+
+    # Get UV coordinates and depth for each vertex
+    uv, depth_vert = get_uv_depth(vertices, depth_fitted, depth_sparse, 10)
+
+    # Backproject to 3D
+    vertices_3d = project_vertices_to_3d(uv, depth_vert, col_img)
+
+    return vertices_3d
+
+
+def merge_vertices_3d(vert_edge_per_image, th=0.5):
+    '''Merge vertices that are close to each other in 3D space and are of same types'''
+    # Initialize structures to collect vertices and connections from all images
+    all_3d_vertices = []
+    connections_3d = []
+    all_indexes = []
+    cur_start = 0
+    types = []
+    
+    # Combine vertices and update connection indices across all images
+    for cimg_idx, (vertices, connections, vertices_3d) in vert_edge_per_image.items():
+        types += [int(v['type']=='apex') for v in vertices]
+        all_3d_vertices.append(vertices_3d)
+        connections_3d+=[(x+cur_start,y+cur_start) for (x,y) in connections]
+        cur_start+=len(vertices_3d)
+    all_3d_vertices = np.concatenate(all_3d_vertices, axis=0)
+    
+    # Calculate distance matrix between all vertices
+    distmat = cdist(all_3d_vertices, all_3d_vertices)
+    types = np.array(types).reshape(-1,1)
+    same_types = cdist(types, types)
+    
+    # Create mask for vertices that should be merged (close in space and same type)
+    mask_to_merge = (distmat <= th) & (same_types==0)
+    new_vertices = []
+    new_connections = []
+    
+    # Extract vertex indices to merge based on the mask
+    to_merge = sorted(list(set([tuple(a.nonzero()[0].tolist()) for a in mask_to_merge])))
+    
+    # Build groups of vertices to merge (transitive grouping)
+    to_merge_final = defaultdict(list)
+    for i in range(len(all_3d_vertices)):
+        for j in to_merge:
+            if i in j:
+                to_merge_final[i]+=j
+    
+    # Remove duplicates in each group
+    for k, v in to_merge_final.items():
+        to_merge_final[k] = list(set(v))
+    
+    # Create final merge groups without duplicates
+    already_there = set() 
+    merged = []
+    for k, v in to_merge_final.items():
+        if k in already_there:
+            continue
+        merged.append(v)
+        for vv in v:
+            already_there.add(vv)
+    
+    # Calculate new vertex positions (average of merged groups)
+    old_idx_to_new = {}
+    count=0
+    for idxs in merged:
+        new_vertices.append(all_3d_vertices[idxs].mean(axis=0))
+        for idx in idxs:
+            old_idx_to_new[idx] = count
+        count +=1
+    new_vertices=np.array(new_vertices)
+    
+    # Update connections to use new vertex indices
+    for conn in connections_3d:
+        new_con = sorted((old_idx_to_new[conn[0]], old_idx_to_new[conn[1]]))
+        if new_con[0] == new_con[1]:
+            continue
+        if new_con not in new_connections:
+            new_connections.append(new_con)
+    return new_vertices, new_connections
+
+
+def prune_not_connected(all_3d_vertices, connections_3d, keep_largest=True):
+    """
+    Prune vertices not connected to anything. If keep_largest=True, also
+    keep only the largest connected component in the graph.
+    """
+    if len(all_3d_vertices) == 0:
+        return np.array([]), []
+
+    # adjacency
+    adj = defaultdict(set)
+    for (i, j) in connections_3d:
+        adj[i].add(j)
+        adj[j].add(i)
+
+    # keep only vertices that appear in at least one edge
+    used_idxs = set()
+    for (i, j) in connections_3d:
+        used_idxs.add(i)
+        used_idxs.add(j)
+
+    if not used_idxs:
+        return np.empty((0,3)), []
+
+    # If we only want to remove truly isolated points, but keep multiple subgraphs:
+    if not keep_largest:
+        new_map = {}
+        used_list = sorted(list(used_idxs))
+        for new_id, old_id in enumerate(used_list):
+            new_map[old_id] = new_id
+        new_vertices = np.array([all_3d_vertices[old_id] for old_id in used_list])
+        new_conns = []
+        for (i, j) in connections_3d:
+            if i in used_idxs and j in used_idxs:
+                new_conns.append((new_map[i], new_map[j]))
+        return new_vertices, new_conns
+
+    # Otherwise find the largest connected component:
+    visited = set()
+    def bfs(start):
+        queue = [start]
+        comp = []
+        visited.add(start)
+        while queue:
+            cur = queue.pop()
+            comp.append(cur)
+            for neigh in adj[cur]:
+                if neigh not in visited:
+                    visited.add(neigh)
+                    queue.append(neigh)
+        return comp
+
+    # Collect all subgraphs
+    comps = []
+    for idx in used_idxs:
+        if idx not in visited:
+            c = bfs(idx)
+            comps.append(c)
+
+    # pick largest
+    comps.sort(key=lambda c: len(c), reverse=True)
+    largest = comps[0] if len(comps)>0 else []
+
+    # Remap
+    new_map = {}
+    for new_id, old_id in enumerate(largest):
+        new_map[old_id] = new_id
+
+    new_vertices = np.array([all_3d_vertices[old_id] for old_id in largest])
+    new_conns = []
+    for (i, j) in connections_3d:
+        if i in largest and j in largest:
+            new_conns.append((new_map[i], new_map[j]))
+
+    # remove duplicates
+    new_conns = list(set([tuple(sorted(c)) for c in new_conns]))
+    return new_vertices, new_conns
+
+def get_sparse_depth(colmap_rec, img_id_substring, depth):
+    """
+    Return a sparse depth map for the COLMAP image whose name contains
+    `img_id_substring`. The output is an array of shape `depth_shape` (H,W),
+    where only the projected 3D points get a depth > 0, else 0.
+    """
+    H, W = depth.shape
+
+    # 1) Find the matching COLMAP image
+    found_img = None
+    for img_id_c, col_img in colmap_rec.images.items():
+        if img_id_substring in col_img.name:
+            found_img = col_img
+            break
+    if found_img is None:
+        print(f"Image substring {img_id_substring} not found in COLMAP.")
+        return np.zeros((H, W), dtype=np.float32), False, None
+    
+    # 2) Gather 3D points that this image sees
+    points_xyz = []
+    for pid, p3D in colmap_rec.points3D.items():
+        if found_img.has_point3D(pid):
+            points_xyz.append(p3D.xyz)  # world coords
+    if not points_xyz:
+        print(f"No 3D points associated with {found_img.name}.")
+        return np.zeros((H, W), dtype=np.float32), False, found_img
+    
+    points_xyz = np.array(points_xyz)  # (N, 3)
+    
+    # 3) For each point, project via col_img.project_point()
+    uv = []
+    z_vals = []
+    for xyz in points_xyz:
+        proj = found_img.project_point(xyz)  # returns (u, v) in image coords or None
+        if proj is not None:
+            u_i, v_i = proj
+            u_i = int(round(u_i))
+            v_i = int(round(v_i))
+            # Check in-bounds
+            if 0 <= u_i < W and 0 <= v_i < H:
+                uv.append((u_i, v_i))
+                # We'll compute depth as Z in camera coords
+                # from the world->cam transform col_img holds
+                mat4x4 = np.eye(4)
+                mat4x4[:3, :4] = found_img.cam_from_world.matrix()
+                p_cam =  mat4x4@ np.array([xyz[0], xyz[1], xyz[2], 1.0])
+                z_vals.append(p_cam[2] / p_cam[3]) 
+    
+    uv = np.array(uv, dtype=int)     # shape (M,2)
+    z_vals = np.array(z_vals)        # shape (M,)
+    
+    depth_out = np.zeros((H, W), dtype=np.float32)
+    depth_out[uv[:,1], uv[:,0]] = z_vals  # Note: uv = (u, v), so row = v, col = u
+    
+    return depth_out, True, found_img
+
+
+def fit_scale_robust_median(depth, sparse_depth, validity_mask=None):
+    """
+    Fit a scale factor to the depth map using the median of the ratio of sparse to dense depth.
+    """
+    if validity_mask is None:
+        mask = (sparse_depth != 0)
+    else:
+        mask = (sparse_depth != 0) & validity_mask
+    mask = mask & (depth <50) & (sparse_depth <50)
+    X = depth[mask]
+    Y = sparse_depth[mask]
+    alpha =np.median(Y/X)
+    depth_fitted = alpha * depth
+    return alpha, depth_fitted
+    
+
+def get_fitted_dense_depth(depth, colmap_rec, img_id, ade20k_seg):
+    """
+    Gets sparse depth from COLMAP, computes a house mask, fits dense depth to sparse 
+    depth within the mask, and returns the fitted dense depth.
+
+    Parameters
+    ----------
+    depth : np.ndarray
+        Initial dense depth map (H, W).
+    colmap_rec : pycolmap.Reconstruction
+        COLMAP reconstruction data.
+    img_id : str
+        Identifier for the current image within the COLMAP reconstruction.
+    K : np.ndarray
+        Camera intrinsic matrix (3x3).
+    R : np.ndarray
+        Camera rotation matrix (3x3).
+    t : np.ndarray
+        Camera translation vector (3,).
+    ade20k_seg : PIL.Image
+        ADE20k segmentation map for the image.
+
+    Returns
+    -------
+    depth_fitted : np.ndarray
+        Dense depth map scaled and shifted to align with sparse depth within the house mask (H, W).
+    depth_sparse : np.ndarray
+        The sparse depth map obtained from COLMAP (H, W).
+    found_sparse : bool
+        True if sparse depth points were found for this image, False otherwise.
+    """
+    depth_np = np.array(depth) / 1000. # Convert mm to meters if needed
+    depth_sparse, found_sparse, col_img = get_sparse_depth(colmap_rec, img_id, depth_np)
+    
+    if not found_sparse:
+        print(f'No sparse depth found for image {img_id}')
+        # Return original (meter-scaled) depth if no sparse data
+        return depth_np, np.zeros_like(depth_np), False, None
+
+    # Get house mask to focus fitting on relevant areas
+    house_mask = get_house_mask(ade20k_seg)
+    
+    # Fit dense depth to sparse depth (scale only), using only points within the house mask
+    k, depth_fitted = fit_scale_robust_median(depth_np, depth_sparse, validity_mask=house_mask)
+    print(f"Fitted depth scale k={k:.4f} for image {img_id}")
+    #depth_fitted = depth_np# * house_mask.astype(np.float32)
+    depth_sparse = depth_sparse# * house_mask.astype(np.float32)
+    return depth_fitted, depth_sparse, True, col_img
+
+
+def prune_too_far(all_3d_vertices, connections_3d, colmap_rec, th = 3.0):
+    """
+    Prune vertices that are too far from sparse point cloud
+    
+    """
+    xyz_sfm=[]
+    for k, v in colmap_rec.points3D.items():
+        xyz_sfm.append(v.xyz)
+    xyz_sfm = np.array(xyz_sfm)
+    distmat = cdist(all_3d_vertices, xyz_sfm)
+    mindist = distmat.min(axis=1)
+    mask = mindist <= th
+    all_3d_vertices_new = all_3d_vertices[mask]
+    old_idx_survived = np.arange(len(all_3d_vertices))[mask]
+    new_idxs = np.arange(len(all_3d_vertices_new))
+    old_to_new_idx = dict(zip(old_idx_survived, new_idxs))
+    connections_3d_new = [(old_to_new_idx[conn[0]], old_to_new_idx[conn[1]]) for conn in connections_3d if mask[conn[0]] and mask[conn[1]]]   
+    return all_3d_vertices_new, connections_3d_new
+
+
+def predict_wireframe(entry) -> Tuple[np.ndarray, List[int]]:
+    """
+    Predict 3D wireframe from a dataset entry.
+    """
+    good_entry = convert_entry_to_human_readable(entry)
+    vert_edge_per_image = {}
+    for i, (gest, depth, K, R, t, img_id, ade_seg) in enumerate(zip(good_entry['gestalt'],
+                                                good_entry['depth'], 
+                                                good_entry['K'],
+                                                good_entry['R'],
+                                                good_entry['t'],
+                                                good_entry['image_ids'],
+                                                good_entry['ade'] # Added ade20k segmentation
+                                                )):
+        colmap_rec = good_entry['colmap_binary']
+        K = np.array(K)
+        R = np.array(R)
+        t = np.array(t)
+        # Resize gestalt segmentation to match depth map size
+        depth_size = (np.array(depth).shape[1], np.array(depth).shape[0]) # W, H
+        gest_seg = gest.resize(depth_size)
+        gest_seg_np = np.array(gest_seg).astype(np.uint8)
+        
+        # Get 2D vertices and edges first
+        vertices, connections = get_vertices_and_edges_from_segmentation(gest_seg_np, edge_th=10.)
+        
+        # Check if we have enough to proceed
+        if (len(vertices) < 2) or (len(connections) < 1):
+            print(f'Not enough vertices or connections found in image {i}, skipping.')
+            vert_edge_per_image[i] = [], [], np.empty((0, 3))
+            continue
+            
+        # Call the refactored function to get 3D points
+        vertices_3d = create_3d_wireframe_single_image(
+            vertices, connections, depth, colmap_rec, img_id, ade_seg
+        )
+        # Store original 2D vertices, connections, and computed 3D points
+        vert_edge_per_image[i] = vertices, connections, vertices_3d
+    
+    # Merge vertices from all images
+    all_3d_vertices, connections_3d = merge_vertices_3d(vert_edge_per_image, 0.5)
+    all_3d_vertices_clean, connections_3d_clean  = prune_not_connected(all_3d_vertices, connections_3d, keep_largest=False)
+    all_3d_vertices_clean, connections_3d_clean  = prune_too_far(all_3d_vertices_clean, connections_3d_clean, colmap_rec, th = 4.0)
+    
+    if (len(all_3d_vertices_clean) < 2) or len(connections_3d_clean) < 1:
+        print (f'Not enough vertices or connections in the 3D vertices')
+        return empty_solution()
+
+    return all_3d_vertices_clean, connections_3d_clean 

{hoho → hoho2025}/hoho.py

@@ -13,6 +13,7 @@ import numpy as np
 import importlib
 import subprocess
 
+
 from PIL import ImageFile
 
 from huggingface_hub.utils._headers import build_hf_headers # note: using _headers
@@ -184,8 +185,9 @@ def proc(row, split='train'):
     return Sample(out)
 
 
-from . import read_write_colmap
+
 def decode_colmap(s):
+    import hoho2025.read_write_colmap as read_write_colmap
     with temp_working_directory():
 
         with open('points3D.bin', 'wb') as stream:

hoho2025/metric_helper.py

@@ -0,0 +1,167 @@
+import numpy as np
+from scipy.spatial.distance import cdist
+from scipy.optimize import linear_sum_assignment
+import torch
+import trimesh
+from time import time
+
+MAX_SCORE = 1.0
+
+def get_one_primitive(p1, p2, c=(255, 0, 0), radius=25, primitive_type='cylinder', sections=6):
+    if len(c) == 1:
+        c = [c[0]] * 4
+    elif len(c) == 3:
+        c = [*c, 255]
+    elif len(c) != 4:
+        raise ValueError(f'{c} is not a valid color (must have 1,3, or 4 elements).')
+
+    p1, p2 = np.asarray(p1), np.asarray(p2)
+    l = np.linalg.norm(p2 - p1)
+    
+    # Add check for zero-length edges
+    if l < 1e-6:
+        return None
+        
+    direction = (p2 - p1) / l
+
+    T = np.eye(4)
+    T[:3, 2] = direction
+    T[:3, 3] = (p1 + p2) / 2
+
+    b0, b1 = T[:3, 0], T[:3, 1]
+    if np.abs(np.dot(b0, direction)) < np.abs(np.dot(b1, direction)):
+        T[:3, 1] = -np.cross(b0, direction)
+    else:
+        T[:3, 0] = np.cross(b1, direction)
+
+    if primitive_type == 'capsule':
+        mesh = trimesh.primitives.Capsule(radius=radius, height=l, transform=T, sections=sections)
+    elif primitive_type == 'cylinder':
+        mesh = trimesh.primitives.Cylinder(radius=radius, height=l, transform=T, sections=sections)
+    else:
+        raise ValueError("Unknown primitive!")
+
+    # Add vertex color initialization check
+    if not hasattr(mesh.visual, 'vertex_colors') or mesh.visual.vertex_colors is None:
+        mesh.visual.vertex_colors = np.ones((len(mesh.vertices), 4)) * 255
+        
+    mesh.visual.vertex_colors = np.ones_like(mesh.visual.vertex_colors) * c
+    return mesh
+
+def get_primitives(vertices, edges, radius=25, c=[255, 0, 0]):
+    # Convert vertices to a NumPy array
+    if isinstance(vertices, torch.Tensor):
+        vertices = vertices.detach().cpu().numpy()
+    else:
+        vertices = np.asarray(vertices)
+    
+    # Convert edges to a NumPy array of integers
+    if isinstance(edges, torch.Tensor):
+        edges = edges.detach().cpu().numpy().astype(np.int64)
+    else:
+        edges = np.asarray(edges, dtype=np.int64)
+    
+    primitives = []
+    for e in edges:
+        # Add edge validation
+        if e[0] >= len(vertices) or e[1] >= len(vertices):
+            continue
+        primitive = get_one_primitive(vertices[e[0]], vertices[e[1]], radius=radius, c=c)
+        if primitive is not None:
+            primitives.append(primitive)
+    return primitives
+
+
+
+def compute_mesh_iou_VOLUME(pd_vertices, pd_edges, gt_vertices, gt_edges, radius=20, engine='manifold'):
+    # check empty 
+    if len(pd_edges) == 0 or len(gt_edges) == 0:
+        return 0.0
+
+    pd_vertices = pd_vertices.detach().cpu() if isinstance(pd_vertices, torch.Tensor) else pd_vertices
+    pd_edges = pd_edges.detach().cpu() if isinstance(pd_edges, torch.Tensor) else pd_edges
+    gt_vertices = gt_vertices.detach().cpu() if isinstance(gt_vertices, torch.Tensor) else gt_vertices
+    gt_edges = gt_edges.detach().cpu() if isinstance(gt_edges, torch.Tensor) else gt_edges
+
+    pd_primitives = get_primitives(pd_vertices, pd_edges, radius=radius, c=[0, 255, 0])
+    gt_primitives = get_primitives(gt_vertices, gt_edges, radius=radius, c=[255, 0, 0])
+    # check for empty primitives
+    if not pd_primitives or not gt_primitives:
+        return 0.0
+
+    # Add bounding box check to detect non-overlapping cases quickly
+    pd_bounds = np.array([p.bounds for p in pd_primitives])
+    gt_bounds = np.array([p.bounds for p in gt_primitives])
+    
+    pd_min, pd_max = np.min(pd_bounds[:, 0], axis=0), np.max(pd_bounds[:, 1], axis=0)
+    gt_min, gt_max = np.min(gt_bounds[:, 0], axis=0), np.max(gt_bounds[:, 1], axis=0)
+    
+    # If bounding boxes don't overlap, return 0
+    if np.any(pd_max < gt_min) or np.any(pd_min > gt_max):
+        return 0.0
+    t=time()
+    mesh_pred = trimesh.boolean.union(pd_primitives, engine=engine)
+    #print(f"mesh_pred union: {time() - t} {mesh_pred.is_volume}")
+    t=time()
+    mesh_gt= trimesh.boolean.union(gt_primitives, engine=engine)
+    #print(f"mesh_gt union: {time() - t} {mesh_gt.is_volume}")
+
+    if mesh_pred.is_volume and mesh_gt.is_volume:
+        t=time()
+        inter_volume = trimesh.boolean.intersection([mesh_pred, mesh_gt], engine=engine).volume
+        #print(f"inter_volume: {time() - t}")
+    else:
+        all_inter = []
+        t=time()
+        for pd_prim in pd_primitives:
+            pd_min, pd_max = pd_prim.bounds
+            for gt_prim in gt_primitives:
+                # Skip intersection calculation if bounding boxes don't overlap
+                gt_min, gt_max = gt_prim.bounds
+                if np.any(pd_max < gt_min) or np.any(pd_min > gt_max):
+                    continue
+                inter = trimesh.boolean.intersection([pd_prim, gt_prim], engine=engine)
+                if inter.is_volume and inter.volume > 0:
+                    all_inter.append(inter)
+        inter_volume = trimesh.boolean.union(all_inter, engine=engine).volume if all_inter else 0
+        #print(f"all_inter: {time() - t}")
+    union_volume = mesh_pred.volume + mesh_gt.volume - inter_volume
+    
+    return inter_volume / union_volume if union_volume > 0 else 0.0
+
+
+# ----------------- Corner F1 -----------------
+def compute_ap_metrics(pd_vertices, gt_vertices, thresh=25):
+    if len(pd_vertices) == 0 or len(gt_vertices) == 0:
+        return 0.0
+
+    dists = cdist(pd_vertices, gt_vertices)
+    row_ind, col_ind = linear_sum_assignment(dists)
+
+    tp = (dists[row_ind, col_ind] <= thresh).sum()
+    precision = tp / len(pd_vertices) if len(pd_vertices) > 0 else 0
+    recall = tp / len(gt_vertices) if len(gt_vertices) > 0 else 0
+    denom = precision + recall
+    f1 = (2 * precision * recall / denom) if denom > 0 else 0.0
+    return f1
+
+def batch_corner_f1(X, Y, distance_thresh=25):
+    results = []
+    for (pd_v, _), (gt_v, _) in zip(X, Y):
+        results.append(compute_ap_metrics(pd_v, gt_v, thresh=distance_thresh))
+    return np.array(results)
+
+# ----------------- HSS Metric -----------------
+from collections import namedtuple
+HSSReturnType = namedtuple('HSSReturnType', ['hss', 'f1', 'iou'])
+def hss(y_hat_v, y_hat_e, y_v, y_e, vert_thresh=0.5, edge_thresh=0.5):
+    X = [(y_hat_v, y_hat_e)]
+    Y = [(y_v, y_e)]
+    t=time()
+    f1 = np.clip(batch_corner_f1(X, Y, distance_thresh=vert_thresh)[0], 0, 1)
+    #print(f"f1 {f1}: in {time() - t:.2f} sec")
+    t=time()
+    IoU = np.clip(compute_mesh_iou_VOLUME(y_hat_v, y_hat_e, y_v, y_e, radius=edge_thresh), 0, 1)
+    #print(f"IoU: {IoU} in {time() - t:.2f} sec")
+    score = 2 * f1 * IoU / (f1 + IoU) if (f1 + IoU) > 0 else 0.0
+    return HSSReturnType(hss=score, f1=f1, iou=IoU)

{hoho → hoho2025}/read_write_colmap.py

@@ -486,4 +486,3 @@ def rotmat2qvec(R):
     if qvec[0] < 0:
         qvec *= -1
     return qvec
-

{hoho → hoho2025}/vis.py

@@ -1,3 +1,5 @@
+
+import matplotlib.pyplot as plt
 import trimesh
 import numpy as np
 from copy import deepcopy
@@ -5,6 +7,39 @@ from PIL import Image
 
 from . import color_mappings
 
+
+def plot_all_modalities(ds_entry, figsize=(8, 15)):
+    modalities_to_plot = ['images', 'depth', 'gestalt', 'ade']
+    modalities_in_entry = [k for k in ds_entry.keys() if k in modalities_to_plot and len(ds_entry[k]) > 0]
+    number_of_columns = len(modalities_in_entry)
+    number_of_images = len(ds_entry['image_ids'])
+    number_of_rows = number_of_images
+    fig, axes = plt.subplots(number_of_rows, number_of_columns, figsize=figsize)
+    for i in range(len(ds_entry[modalities_in_entry[0]])):
+        for j, modality in enumerate(modalities_in_entry):
+            ax = axes[i, j]
+            if modality == 'image':
+                ax.imshow(ds_entry[modality][i])
+            elif modality == 'depth':
+                depth_image = np.array(ds_entry[modality][i])/1000.0
+                ax.imshow(depth_image, cmap='rainbow')
+            elif modality == 'gestalt':
+                ax.imshow(ds_entry[modality][i])
+            elif modality == 'ade':
+                ax.imshow(ds_entry[modality][i])
+            else:
+                raise ValueError(f"Unknown modality: {modality}")
+            if i == 0:
+                ax.set_title(modality)
+            ax.axis('off')
+            if j == 0:
+                ax.set_ylabel(f"Image {i}")
+    fig.tight_layout()  
+    fig.subplots_adjust(wspace=0.05, hspace=0.01)
+    #plt.show()
+    return fig, axes
+
+
 def line(p1, p2, c=(255,0,0), resolution=10, radius=0.05):
     '''draws a 3d cylinder along the line (p1, p2)'''
     # check colors
@@ -114,8 +149,6 @@ def show_grid(edges, meshes=None, row_length=5):
     return trimesh.Scene(out)
 
 
-
-
 def visualize_order_images(row_order):
     return create_image_grid(row_order['ade20k'] + row_order['gestalt'] + [visualize_depth(dm) for dm in row_order['depthcm']], num_per_row=len(row_order['ade20k']))
 
@@ -146,8 +179,6 @@ def create_image_grid(images, target_length=312, num_per_row=2):
     return grid_img
 
 
-import matplotlib.pyplot as plt
-
 def visualize_depth(depth, min_depth=None, max_depth=None, cmap='rainbow'):
     depth = np.array(depth)
     

{hoho → hoho2025}/viz3d.py

@@ -1,4 +1,3 @@
-
 """
 Copyright [2022] [Paul-Edouard Sarlin and Philipp Lindenberger]
 
@@ -21,58 +20,23 @@ Works for a small number of points and cameras, might be slow otherwise.
 2) Add 3D points, camera frustums, or both as a pycolmap.Reconstruction
 
 Written by Paul-Edouard Sarlin and Philipp Lindenberger.
+Slightly modified by Dmytro Mishkin
 """
-# Slightly modified by Dmytro Mishkin
-
 from typing import Optional
 import numpy as np
 import pycolmap
 import plotly.graph_objects as go
-
-
-### Some helper functions for geometry
-def qvec2rotmat(qvec):
-    return np.array([
-        [1 - 2 * qvec[2]**2 - 2 * qvec[3]**2,
-         2 * qvec[1] * qvec[2] - 2 * qvec[0] * qvec[3],
-         2 * qvec[3] * qvec[1] + 2 * qvec[0] * qvec[2]],
-        [2 * qvec[1] * qvec[2] + 2 * qvec[0] * qvec[3],
-         1 - 2 * qvec[1]**2 - 2 * qvec[3]**2,
-         2 * qvec[2] * qvec[3] - 2 * qvec[0] * qvec[1]],
-        [2 * qvec[3] * qvec[1] - 2 * qvec[0] * qvec[2],
-         2 * qvec[2] * qvec[3] + 2 * qvec[0] * qvec[1],
-         1 - 2 * qvec[1]**2 - 2 * qvec[2]**2]])
-
+from hoho2025.color_mappings import edge_color_mapping, EDGE_CLASSES_BY_ID
 
 def to_homogeneous(points):
     pad = np.ones((points.shape[:-1]+(1,)), dtype=points.dtype)
     return np.concatenate([points, pad], axis=-1)
 
-def t_to_proj_center(qvec, tvec):
-    Rr = qvec2rotmat(qvec)
-    tt = (-Rr.T) @ tvec
-    return tt
-
-def calib(params):
-    out = np.eye(3)
-    if len(params) == 3:
-        out[0,0] = params[0]
-        out[1,1] = params[0]
-        out[0,2] = params[1]
-        out[1,2] = params[2]
-    else:
-        out[0,0] = params[0]
-        out[1,1] = params[1]
-        out[0,2] = params[2]
-        out[1,2] = params[3]
-    return out
-
-
 ### Plotting functions
 
 def init_figure(height: int = 800) -> go.Figure:
     """Initialize a 3D figure."""
-    fig = go.Figure()
+    fig = go.FigureWidget()
     axes = dict(
         visible=False,
         showbackground=False,
@@ -118,9 +82,14 @@ def plot_lines_3d(
     x = pts[..., 0]
     y = pts[..., 1]
     z = pts[..., 2]
-    traces = [go.Scatter3d(x=x1, y=y1, z=z1,
+    if isinstance(color, list):
+        traces = [go.Scatter3d(x=x1, y=y1, z=z1,
+                            mode='lines',
+                            line=dict(color=f"rgb{c}", width=ps)) for x1, y1, z1, c in zip(x,y,z,color)]
+    else:
+        traces = [go.Scatter3d(x=x1, y=y1, z=z1,
                         mode='lines',
-                        line=dict(color=color, width=2)) for x1, y1, z1 in zip(x,y,z)]
+                        line=dict(color=color, width=ps)) for x1, y1, z1 in zip(x,y,z)]
     for t in traces:
         fig.add_trace(t)
     fig.update_traces(showlegend=False)
@@ -150,7 +119,11 @@ def plot_camera(
         name: Optional[str] = None,
         legendgroup: Optional[str] = None,
         size: float = 1.0):
-    """Plot a camera frustum from pose and intrinsic matrix."""
+    """Plot a camera frustum from pose and intrinsic matrix. R and t are 
+    world_to_camera transformation"""
+    R = np.array(R)
+    t = np.array(t).reshape(3)
+    K = np.array(K)
     W, H = K[0, 2]*2, K[1, 2]*2
     corners = np.array([[0, 0], [W, 0], [W, H], [0, H], [0, 0]])
     if size is not None:
@@ -197,106 +170,118 @@ def plot_camera_colmap(
         name: Optional[str] = None,
         **kwargs):
     """Plot a camera frustum from PyCOLMAP objects"""
-    intr = calib(camera.params)
-    if intr[0][0] > 10000:
+    # Use camera intrinsics method if available, otherwise fallback to params
+    intr = camera.calibration_matrix()
+    if intr[0][0] > 5000:
         print("Bad camera")
         return
+    world_t_camera = image.cam_from_world.inverse()
     plot_camera(
         fig,
-        qvec2rotmat(image.qvec).T,
-        t_to_proj_center(image.qvec, image.tvec),
-        intr,#calibration_matrix(),
-        name=name or str(image.id),
+        world_t_camera.rotation.matrix(),  # Use rotation matrix method (World-to-Camera)
+        world_t_camera.translation,  # Use camera center in world coordinates
+        intr,
+        name=name or str(image.name),
         **kwargs)
 
 
 def plot_cameras(
         fig: go.Figure,
-        reconstruction,#: pycolmap.Reconstruction,
+        reconstruction: pycolmap.Reconstruction, # Added type hint
         **kwargs):
     """Plot a camera as a cone with camera frustum."""
-    for image_id, image in reconstruction["images"].items():
+    # Iterate over reconstruction.images
+    for image_id, image in reconstruction.images.items():
+        # Access camera using reconstruction.cameras
         plot_camera_colmap(
-            fig, image, reconstruction["cameras"][image.camera_id], **kwargs)
+            fig, image, reconstruction.cameras[image.camera_id], **kwargs)
 
 
 def plot_reconstruction(
         fig: go.Figure,
-        rec,
+        rec: pycolmap.Reconstruction, # Added type hint
         color: str = 'rgb(0, 0, 255)',
         name: Optional[str] = None,
         points: bool = True,
         cameras: bool = True,
         cs: float = 1.0,
         single_color_points=False,
-        camera_color='rgba(0, 255, 0, 0.5)'):
-    # rec is result of loading reconstruction from "read_write_colmap.py"
+        camera_color='rgba(0, 255, 0, 0.5)',
+        crop_outliers: bool = False):
+    # rec is a pycolmap.Reconstruction object
     # Filter outliers
     xyzs = []
     rgbs = []
-    for k, p3D in rec['points'].items():
+    # Iterate over rec.points3D
+    for k, p3D in rec.points3D.items():
+        #print (p3D)
         xyzs.append(p3D.xyz)
-        rgbs.append(p3D.rgb)
-
-    if points:
-        plot_points(fig, np.array(xyzs), color=color if single_color_points else np.array(rgbs), ps=1, name=name)
+        rgbs.append(p3D.color)
+    
+    xyzs = np.array(xyzs)
+    rgbs = np.array(rgbs)
+    
+    # Crop outliers if requested
+    if crop_outliers and len(xyzs) > 0:
+        # Calculate distances from origin
+        distances = np.linalg.norm(xyzs, axis=1)
+        # Find threshold at 98th percentile (removing 2% furthest points)
+        threshold = np.percentile(distances, 98)
+        # Filter points
+        mask = distances <= threshold
+        xyzs = xyzs[mask]
+        rgbs = rgbs[mask]
+        print(f"Cropped outliers: removed {np.sum(~mask)} out of {len(mask)} points ({np.sum(~mask)/len(mask)*100:.2f}%)")
+
+    if points and len(xyzs) > 0:
+        plot_points(fig, xyzs, color=color if single_color_points else rgbs, ps=1, name=name)
     if cameras:
         plot_cameras(fig, rec, color=camera_color, legendgroup=name, size=cs)
 
-
-def plot_pointcloud(
+def plot_wireframe(
         fig: go.Figure,
-        pts: np.ndarray,
-        colors: np.ndarray,
-        ps: int = 2,
-        name: Optional[str] = None):
-    """Plot a set of 3D points."""
-    plot_points(fig, np.array(pts), color=colors, ps=ps, name=name)
-
-
-def plot_triangle_mesh(
-        fig: go.Figure,
-        vert: np.ndarray,
-        colors: np.ndarray,
-        triangles: np.ndarray,
-        name: Optional[str] = None):
-    """Plot a triangle mesh."""
-    tr = go.Mesh3d(
-        x=vert[:,0],
-        y=vert[:,1],
-        z=vert[:,2],
-        vertexcolor = np.clip(255*colors, 0, 255),
-        # i, j and k give the vertices of triangles
-        # here we represent the 4 triangles of the tetrahedron surface
-        i=triangles[:,0],
-        j=triangles[:,1],
-        k=triangles[:,2],
-        name=name,
-        showscale=False
-    )
-    fig.add_trace(tr)
-
-def plot_estimate_and_gt(pred_vertices, pred_connections, gt_vertices=None, gt_connections=None):
-    fig3d = init_figure()
-    c1 = (30, 20, 255)
-    img_color = [c1 for _ in range(len(pred_vertices))]
-    plot_points(fig3d, pred_vertices, color = img_color, ps = 10)  
-    lines = []
-    for c in pred_connections:
-        v1 = pred_vertices[c[0]]
-        v2 = pred_vertices[c[1]]
-        lines.append(np.stack([v1, v2], axis=0))
-    plot_lines_3d(fig3d, np.array(lines), img_color, ps=4)  
+        vertices: np.ndarray,
+        edges: np.ndarray,
+        classifications: np.ndarray = None,
+        color: str = 'rgb(0, 0, 255)',
+        name: Optional[str] = None,
+        **kwargs):
+    """Plot a camera as a cone with camera frustum."""
+    gt_vertices = np.array(vertices)
+    gt_connections = np.array(edges)
     if gt_vertices is not None:
-        c2 = (30, 255, 20)
-        img_color2 = [c2 for _ in range(len(gt_vertices))]
-        plot_points(fig3d, gt_vertices, color = img_color2, ps = 10)  
+        img_color2 = [color for _ in range(len(gt_vertices))]
+        plot_points(fig, gt_vertices, color = img_color2, ps = 10)  
         if gt_connections is not None:
             gt_lines = []
             for c in gt_connections:
                 v1 = gt_vertices[c[0]]
                 v2 = gt_vertices[c[1]]
                 gt_lines.append(np.stack([v1, v2], axis=0))
-        plot_lines_3d(fig3d, np.array(gt_lines), img_color2, ps=4)        
-    fig3d.show()
-    return fig3d
+            if classifications is not None and len(classifications) == len(gt_lines):
+                line_colors = []
+                for c in classifications:
+                    line_colors.append(edge_color_mapping[EDGE_CLASSES_BY_ID[c]])
+                plot_lines_3d(fig, np.array(gt_lines), line_colors, ps=4)  
+            else:
+                plot_lines_3d(fig, np.array(gt_lines), color, ps=4)  
+
+
+def plot_bpo_cameras_from_entry(fig: go.Figure, entry: dict, idx = None):
+    def cam2world_to_world2cam(R, t):
+        rt = np.eye(4)
+        rt[:3,:3] = R
+        rt[:3,3] = t.reshape(-1)
+        rt = np.linalg.inv(rt)
+        return rt[:3,:3], rt[:3,3]
+    
+    for i in range(len(entry['R'])):
+        if idx is not None and i != idx:
+            continue
+        K = np.array(entry['K'][i])
+        R = np.array(entry['R'][i])
+        t = np.array(entry['t'][i])
+        R, t = cam2world_to_world2cam(R, t)
+        plot_camera(fig, R, t, K)
+    
+    

requirements.txt

@@ -1,10 +1,14 @@
 datasets
+huggingface-hub
 ipywidgets
 matplotlib
 numpy
-pillow
+opencv-python
+Pillow
 plotly
 pycolmap
 scipy
+torch
 trimesh
-webdataset
+webdataset
+manifold3d # for metric computation

setup.py

@@ -5,12 +5,13 @@ import glob
 with open('requirements.txt') as f:
     required = f.read().splitlines()
 
-setup(name='hoho',
-	version='0.0.4',
+setup(name='hoho2025',
+	version='0.1.0',
 	description='Tools and utilites for the HoHo Dataset and S23DR Competition',
 	url='usm3d.github.io',
 	author='Jack Langerman, Dmytro Mishkin, S23DR Orgainizing Team',
 	author_email='hoho@jackml.com',
 	install_requires=required,
 	packages=find_packages(),
+	python_requires='>=3.10',
 	include_package_data=True)