def render_parsed_house_walls(parsed_dir, show_pcl=SHOW_PCL, show_by_class=0):
print(f'parsed_dir:{parsed_dir}')
house_name = os.path.basename(parsed_dir)
bboxes = []
labels = []
for obj in CLASSES:
bbox_fn_ = f'{parsed_dir}/object_bbox/{obj}.txt'
bboxes_ = np.loadtxt(bbox_fn_).reshape([-1, 7])
bboxes.append(bboxes_)
label = DSET_METAS0.class_2_label[obj]
labels += [label] * bboxes_.shape[0]
bboxes = np.concatenate(bboxes, 0)
labels = np.array(labels).astype(np.int8)
if bboxes.shape[0] > 0:
scene_size = Bbox3D.boxes_size(bboxes)
print(f'scene wall size:{scene_size}')
#Bbox3D.draw_bboxes(bboxes, up_axis='Z', is_yx_zb=False, labels=labels)
#if not show_pcl:
#Bbox3D.draw_bboxes_mesh(bboxes, up_axis='Z', is_yx_zb=False)
#Bbox3D.draw_bboxes_mesh(bboxes, up_axis='Z', is_yx_zb=False, labels=labels)
#show_walls_offsetz(bboxes)
pass
if show_by_class:
for c in range(1, max(labels) + 1):
cs = DSET_METAS0.label_2_class[c]
print(cs)
if cs not in ['wall', 'window', 'door']:
#if cs not in ['wall']:
continue
mask = labels == c
bboxes_c = bboxes[mask]
show_walls_offsetz(bboxes_c)
if show_pcl:
pcl_fn = f'{parsed_dir}/pcl_camref.ply'
if not os.path.exists(pcl_fn):
return
pcd = open3d.io.read_point_cloud(pcl_fn)
points = np.asarray(pcd.points)
points = cam2world_pcl(points)
colors = np.asarray(pcd.colors)
pcl = np.concatenate([points, colors], 1)
scene_size = pcl_size(pcl)
print(f'scene pcl size:{scene_size}')
print(f'point num: {pcl.shape[0]}')
#pcl = cut_points_roof(pcl)
Bbox3D.draw_points(pcl,
points_keep_rate=POINTS_KEEP_RATE,
points_sample_rate=0.05)
#Bbox3D.draw_points(pcl, points_keep_rate=POINTS_KEEP_RATE, animation_fn='points.mp4', ani_size=AniSizes[house_name])
import pdb
pdb.set_trace() # XXX BREAKPOINT
bboxes[:, 2] += 0.1
def render_pth_file(pth_fn, show_by_class=0):
pcl, bboxes0 = torch.load(pth_fn)
#points = pcl[:,0:3]
#colors = pcl[:,3:6]
#normals = pcl[:,6:9]
bboxes = {}
for c in CLASSES:
bboxes[c] = bboxes0[c]
scene_size = pcl_size(pcl)
print(f'scene pcl size:{scene_size}')
print(f'point num: {pcl.shape[0]}')
#pcl = cut_points_roof(pcl)
classes = [k for k in bboxes.keys()]
num_classes = {k: bboxes[k].shape[0] for k in bboxes.keys()}
print(f'\nclasses: {num_classes}\n\n')
all_bboxes = np.concatenate([boxes for boxes in bboxes.values()], 0)
nums = [boxes.shape[0] for boxes in bboxes.values()]
labels = []
for i, n in enumerate(nums):
labels += [i] * n
labels = np.array(labels)
Bbox3D.draw_points(pcl, points_keep_rate=POINTS_KEEP_RATE)
#show_walls_offsetz(all_bboxes)
#Bbox3D.draw_bboxes_mesh(all_bboxes, up_axis='Z', is_yx_zb=False, labels=labels)
#Bbox3D.draw_bboxes_mesh(all_bboxes, up_axis='Z', is_yx_zb=False)
#Bbox3D.draw_points_bboxes_mesh(pcl, all_bboxes, up_axis='Z', is_yx_zb=False, labels=labels, points_keep_rate=POINTS_KEEP_RATE)
#Bbox3D.draw_points_bboxes_mesh(pcl, all_bboxes, up_axis='Z', is_yx_zb=False, points_keep_rate=POINTS_KEEP_RATE)
#Bbox3D.draw_points_bboxes(pcl, all_bboxes, up_axis='Z', is_yx_zb=False,points_keep_rate=POINTS_KEEP_RATE)
#Bbox3D.draw_points_bboxes(pcl, all_bboxes, up_axis='Z', is_yx_zb=False, labels=labels, points_keep_rate=POINTS_KEEP_RATE)
#Bbox3D.draw_points_bboxes(pcl, all_bboxes, up_axis='Z', is_yx_zb=False, labels=labels, points_keep_rate=POINTS_KEEP_RATE, animation_fn='anima.mp4', ani_size=[280,700,550,1350])
if show_by_class:
for clas in bboxes.keys():
if clas not in ['ceiling', 'floor']:
#if clas not in ['wall', 'window', 'door','ceiling', 'floor']:
continue
print(clas)
#if clas not in CLASSES:
# continue
boxes = bboxes[clas]
#Bbox3D.draw_points_bboxes(points, boxes, up_axis='Z', is_yx_zb=False)
Bbox3D.draw_points_bboxes_mesh(pcl,
boxes,
up_axis='Z',
is_yx_zb=False,
points_keep_rate=POINTS_KEEP_RATE)
show_walls_offsetz(boxes)
import pdb
pdb.set_trace() # XXX BREAKPOINT
pass
def show_pred(gt_boxlists_, pred_boxlists_, files):
SHOW_SMALL_IOU = False
print('SHOW_PRED')
for i in range(len(pred_boxlists_)):
print(f'\n{files[i]}\n')
#pcl_i = dataset[image_ids[i]]['x'][1][:,0:6]
pcl_i = torch.load(files[i])[0][:, 0:6]
preds = pred_boxlists_[i].remove_low('scores', 0.1)
#preds = pred_boxlists_[i] # already post processed in:
select_ids = 0
if select_ids:
ids = [1, 2, 3]
#ids = [4]
preds = preds.select_by_labels(ids, 'labels_org')
gt_boxlists_[i] = gt_boxlists_[i].select_by_labels(
ids, 'labels_org')
if MERGE_BY_CORNERS:
preds = merge_by_corners(preds)
# ~/Research/Detection_3D/maskrcnn_benchmark/modeling/roi_heads/box_head_3d/inference.py
# cfg.MODEL.ROI_HEADS.SCORE_THRESH
xyz_max = pcl_i[:, 0:3].max(0)
xyz_min = pcl_i[:, 0:3].min(0)
xyz_size = xyz_max - xyz_min
pcl_i[:, 0:3] -= xyz_min.reshape([1, 3])
print(f'xyz_size:{xyz_size}')
#preds.show__together(gt_boxlists_[i], points=None, offset_x=xyz_size[0]+0.3, twolabels=False)
#preds.show__together(gt_boxlists_[i], points=pcl_i, offset_x=xyz_size[0]+2.2, twolabels=False, mesh=0, points_keep_rate=0.9, points_sample_rate=1.0, random_color=False)
compare_instances_with_offset = True
if compare_instances_with_offset:
gt_ids = preds.get_field('gt_ids').cpu().data.numpy().astype(
np.int) + 1
if gt_ids.size > 0 and select_ids:
base = np.min(gt_ids[gt_ids > 0]) - 1
gt_ids[gt_ids > 0] -= base
pred_colors = COLOR_LIST[gt_ids].copy()
gt_colors = COLOR_LIST[1:len(gt_boxlists_[i]) + 1].copy()
err_gt_ids = torch.nonzero(gt_boxlists_[i].get_field('labels') ==
0)[:, 0].data.numpy().reshape([-1])
#gt_colors[err_gt_ids] = COLOR_LIST[0].copy()
#preds.show(points=pcl_i, points_keep_rate=0.9, points_sample_rate=1.0, colors=pred_colors)
#gt_boxlists_[i].show(points=pcl_i, points_keep_rate=0.9, points_sample_rate=1.0, colors=gt_colors)
preds.show__together(gt_boxlists_[i],
points=pcl_i,
offset_x=xyz_size[0] + 7,
twolabels=False,
mesh=1,
points_keep_rate=0.9,
points_sample_rate=0.0,
colors=[pred_colors, gt_colors])
#preds.show__together(gt_boxlists_[i],offset_x=xyz_size[0]+7, twolabels=False, mesh=0, points_keep_rate=0.9, points_sample_rate=1.0, colors=[pred_colors, gt_colors])
preds.show__together(gt_boxlists_[i],
points=pcl_i,
offset_x=xyz_size[0] + 7,
twolabels=False,
mesh=0,
points_keep_rate=0.9,
points_sample_rate=0.0,
colors=[pred_colors, gt_colors])
#preds.show_pcl_corners(pcl_i)
Bbox3D.draw_points(pcl_i,
points_keep_rate=0.9,
points_sample_rate=0.6)
import pdb
pdb.set_trace() # XXX BREAKPOINT
pass
#preds.show__together(gt_boxlists_[i], points=pcl_i, offset_x=0, twolabels=True, mesh=0, points_keep_rate=0.9, points_sample_rate=1.0)
#p_labels_org = preds.get_field('labels_org')
#g_labels_org = gt_boxlists_[i].get_field('labels_org')
#if g_labels_org.max() == 5:
# wwd_mask_p = torch.nonzero( p_labels_org <= 3).squeeze()
# wwd_mask_g = torch.nonzero( g_labels_org <= 3).squeeze()
# cf_mask_p = torch.nonzero( p_labels_org > 3).squeeze()
# cf_mask_g = torch.nonzero( g_labels_org > 3).squeeze()
# preds[wwd_mask_p].show__together(gt_boxlists_[i][wwd_mask_g], points=pcl_i, offset_x=xyz_size[0]+2.2, twolabels=False, mesh=False, points_keep_rate=0.9, points_sample_rate=1.0)
# preds[cf_mask_p].show__together(gt_boxlists_[i][cf_mask_g], points=pcl_i, offset_x=xyz_size[0]+2.2, twolabels=False, mesh=False, points_keep_rate=0.9, points_sample_rate=1.0)
#gt_boxlists_[i].show_by_labels([1])
if SHOW_SMALL_IOU:
small_iou_pred_ids = [p['pred_idx'] for p in small_iou_preds[i]]
small_ious = [p['iou'] for p in small_iou_preds[i]]
print(f'small iou preds: {small_iou_preds[i]}')
if len(small_iou_pred_ids) > 0:
preds.show_highlight(small_iou_pred_ids, points=pcl_i)