def save_detections(self, detections, tags, calibs, save_dir):
res_dir = save_dir
if os.path.isdir(res_dir):
shutil.rmtree(res_dir, ignore_errors=True)
os.makedirs(res_dir)
for det, tag, calib in zip(detections, tags, calibs):
label = KittiLabel()
label.current_frame = "Cam2"
final_box_preds = det["box3d_lidar"].detach().cpu().numpy()
label_preds = det["label_preds"].detach().cpu().numpy()
scores = det["scores"].detach().cpu().numpy()
for i in range(final_box_preds.shape[0]):
obj_np = final_box_preds[i, :]
bcenter_Flidar = obj_np[:3].reshape(1, 3)
bcenter_Fcam = calib.lidar2leftcam(bcenter_Flidar)
wlh = obj_np[3:6]
ry = obj_np[-1]
obj = KittiObj()
obj.type = self._class_names[int(label_preds[i])]
obj.score = scores[i]
obj.x, obj.y, obj.z = bcenter_Fcam.flatten()
obj.w, obj.l, obj.h = wlh.flatten()
obj.ry = ry
obj.from_corners(calib, obj.get_bbox3dcorners(), obj.type,
obj.score)
obj.truncated = 0
obj.occluded = 0
obj.alpha = -np.arctan2(-bcenter_Flidar[0, 1],
bcenter_Flidar[0, 0]) + ry
label.add_obj(obj)
# save label
write_str_to_file(str(label), os.path.join(res_dir, f"{tag}.txt"))
def main(is_val, data_dir, vis_dir, result_path):
dt_data = pickle.load(open(result_path, 'rb'))
save_dir = vis_dir
if is_val:
for dt in tqdm(dt_data):
idx = dt['metadata']['image_idx']
idx = "{:06d}".format(idx)
calib, img, label, pc = KittiData(root_dir=data_dir,
idx=idx).read_data()
label_est = KittiLabel()
label_est.data = []
num_dt = len(dt['name'])
for i in range(num_dt):
obj = KittiObj()
obj.type = dt['name'][i]
obj.truncated = dt['truncated'][i]
obj.occluded = dt['occluded'][i]
obj.alpha = dt['alpha'][i]
obj.bbox_l, obj.bbox_t, obj.bbox_r, obj.bbox_b = dt['bbox'][i]
obj.l, obj.h, obj.w = dt['dimensions'][i]
obj.x, obj.y, obj.z = dt['location'][i]
obj.ry = dt['rotation_y'][i]
obj.score = dt['score'][i]
label_est.data.append(obj)
fvimg = FVImage()
fvimg.from_image(img)
for obj in label.data:
if (obj.x > obj.z) or (-obj.x > obj.z):
continue
if obj.type in ['Car', 'Van']:
fvimg.draw_3dbox(obj, calib, bool_gt=True, width=3)
for obj in label_est.data:
if (obj.x > obj.z) or (-obj.x > obj.z):
continue
fvimg.draw_3dbox(obj,
calib,
bool_gt=False,
width=2,
c=(255, 255, int(255 * obj.score), 255))
fvimg_img = Image.fromarray(fvimg.data)
fvimg_img.save(os.path.join(save_dir, "fv_{}.png".format(idx)))
else:
pc_dir = os.path.join(data_dir, "velodyne_points", "data")
img2_dir = os.path.join(data_dir, "image_02", "data")
calib_dir = os.path.join(data_dir, "calib")
calib = read_calib(calib_dir)
for dt in tqdm(dt_data):
idx = dt['metadata']['image_idx']
idx = "{:010d}".format(idx)
pc = read_pc_from_bin(os.path.join(pc_dir, idx + ".bin"))
img = read_image(os.path.join(img2_dir, idx + ".png"))
label_est = KittiLabel()
label_est.data = []
num_dt = len(dt['name'])
for i in range(num_dt):
obj = KittiObj()
obj.type = dt['name'][i]
obj.truncated = dt['truncated'][i]
obj.occluded = dt['occluded'][i]
obj.alpha = dt['alpha'][i]
obj.bbox_l, obj.bbox_t, obj.bbox_r, obj.bbox_b = dt['bbox'][i]
obj.l, obj.h, obj.w = dt['dimensions'][i]
obj.x, obj.y, obj.z = dt['location'][i]
obj.ry = dt['rotation_y'][i]
obj.score = dt['score'][i]
label_est.data.append(obj)
fvimg = FVImage()
fvimg.from_image(img)
# for obj in label.data:
# if obj.type in ['Car', 'Van']:
# fvimg.draw_3dbox(obj, calib, bool_gt=True, width=3)
for obj in label_est.data:
if (obj.x > obj.z) or (-obj.x > obj.z):
continue
fvimg.draw_3dbox(obj,
calib,
bool_gt=False,
width=2,
c=(255, 255, int(obj.score * 255), 255))
fvimg_img = Image.fromarray(fvimg.data)
fvimg_img.save(os.path.join(save_dir, "fv_{}.png".format(idx)))
def test_filt_label_by_range(self):
def limit_period_torch(val, offset=0.5, period=np.pi):
return val - torch.floor(val / period + offset) * period
data_cfg = Test_prep_infos.TRAINDATA_cfg
voxelizer = voxelizer_builder.build(Test_exclude_classes.VOXELIZER_cfg)
target_assigner = target_assigner_builder.build(Test_exclude_classes.TARGETASSIGNER_cfg)
dataloader = build(data_cfg,
ext_dict={
"voxelizer": voxelizer,
"target_assigner": target_assigner,
"feature_map_size": [1, 200, 176]
})
box_coder = target_assigner.box_coder
import torch
import torch.nn as nn
from det3.methods.second.ops.torch_ops import rotate_nms
from det3.dataloader.carladata import CarlaObj, CarlaLabel
from incdet3.utils.utils import filt_label_by_range
for data in dataloader:
tag = data["metadata"][0]["tag"]
if tag != "000006":
continue
label = data["metadata"][0]["label"]
cls_pred = torch.from_numpy(data["labels"]).cuda().float()
cls_pred *= (cls_pred >= 0).float()
cls_pred = cls_pred.long()
cls_pred = nn.functional.one_hot(cls_pred, num_classes=2+1)
cls_pred = cls_pred[..., 1:]
anchors = torch.from_numpy(data["anchors"]).cuda().float()
box_pred = torch.from_numpy(data["reg_targets"]).cuda().float()
box_pred = box_coder.decode(box_pred, anchors)
for box_preds, cls_preds in zip(box_pred, cls_pred):
box_preds = box_preds.float()
cls_preds = cls_preds.float()
total_scores = cls_preds
nms_func = rotate_nms
top_scores, top_labels = torch.max(
total_scores, dim=-1)
top_scores_keep = top_scores >= 0.5
top_scores = top_scores.masked_select(top_scores_keep)
box_preds = box_preds[top_scores_keep]
top_labels = top_labels[top_scores_keep]
boxes_for_nms = box_preds[:, [0, 1, 3, 4, 6]]
selected = nms_func(
boxes_for_nms,
top_scores,
pre_max_size=1000,
post_max_size=1000,
iou_threshold=0.3,
)
selected_boxes = box_preds[selected]
selected_labels = top_labels[selected]
selected_scores = top_scores[selected]
box_preds = selected_boxes
scores = selected_scores
label_preds = selected_labels
final_box_preds = box_preds
final_scores = scores
final_labels = label_preds
predictions_dict = {
"box3d_lidar": final_box_preds,
"scores": final_scores,
"label_preds": label_preds,
}
label_est = KittiLabel()
calib = KittiCalib(f"unit_tests/data/test_kittidata/training/calib/{tag}.txt").read_calib_file()
for box3d_lidar, label_preds, score in zip(
predictions_dict["box3d_lidar"],
predictions_dict["label_preds"],
predictions_dict["scores"]):
obj = KittiObj()
obj.type = "Car" if label_preds == 0 else "Pedestrian"
xyzwlhry_Flidar = box3d_lidar.cpu().numpy().flatten()
bcenter_Flidar = xyzwlhry_Flidar[:3].reshape(-1, 3)
bcenter_Fcam = calib.lidar2leftcam(bcenter_Flidar)
obj.x, obj.y, obj.z = bcenter_Fcam.flatten()
obj.w, obj.l, obj.h, obj.ry = xyzwlhry_Flidar[3:]
obj.truncated = 0
obj.occluded = 0
obj.alpha = 0
obj.bbox_l = 0
obj.bbox_t = 0
obj.bbox_r = 0
obj.bbox_b = 0
label_est.add_obj(obj)
label_est.current_frame = "Cam2"
label_filt = filt_label_by_range(label_est, valid_range=[20, -35.2, -3, 52.8, 12.85, -0.26], calib=calib)
self.assertTrue(len(label_filt) == 1)
def test_targets(self):
def limit_period_torch(val, offset=0.5, period=np.pi):
return val - torch.floor(val / period + offset) * period
import torch
import torch.nn as nn
from det3.methods.second.ops.torch_ops import rotate_nms
from det3.dataloader.kittidata import KittiCalib
for i, data in enumerate(self.dataloader):
if i == 2:
break
else:
continue
label = data["metadata"][0]["label"]
tag = data["metadata"][0]["tag"]
cls_pred = torch.from_numpy(data["labels"]).cuda().float()
cls_pred *= (cls_pred >= 0).float()
cls_pred = cls_pred.long()
cls_pred = nn.functional.one_hot(cls_pred, num_classes=2+1)
cls_pred = cls_pred[..., 1:]
anchors = torch.from_numpy(data["anchors"]).cuda().float()
box_pred = torch.from_numpy(data["reg_targets"]).cuda().float()
# pred_dict = {
# "cls_preds": cls_pred * 10,
# "box_preds": box_pred
# }
# box_coder = self.box_coder
# from det3.ops import write_pkl
# write_pkl({"pred_dict": pred_dict, "box_coder": box_coder}, "test_model_est.pkl")
box_pred = self.box_coder.decode(box_pred, anchors)
for box_preds, cls_preds in zip(box_pred, cls_pred):
box_preds = box_preds.float()
cls_preds = cls_preds.float()
total_scores = cls_preds
nms_func = rotate_nms
top_scores, top_labels = torch.max(
total_scores, dim=-1)
top_scores_keep = top_scores >= 0.5
top_scores = top_scores.masked_select(top_scores_keep)
box_preds = box_preds[top_scores_keep]
top_labels = top_labels[top_scores_keep]
boxes_for_nms = box_preds[:, [0, 1, 3, 4, 6]]
selected = nms_func(
boxes_for_nms,
top_scores,
pre_max_size=1000,
post_max_size=1000,
iou_threshold=0.3,
)
selected_boxes = box_preds[selected]
selected_labels = top_labels[selected]
selected_scores = top_scores[selected]
box_preds = selected_boxes
scores = selected_scores
label_preds = selected_labels
final_box_preds = box_preds
final_scores = scores
final_labels = label_preds
predictions_dict = {
"box3d_lidar": final_box_preds,
"scores": final_scores,
"label_preds": label_preds,
}
from det3.dataloader.kittidata import KittiObj, KittiLabel
label_gt = KittiLabel()
label_est = KittiLabel()
calib = KittiCalib(f"unit_tests/data/test_kittidata/training/calib/{tag}.txt").read_calib_file()
for obj_str in label.split("\n"):
if len(obj_str) == 0:
continue
obj = KittiObj(obj_str)
if obj.type not in ["Car", "Pedestrian"]:
continue
bcenter_Fcam = np.array([obj.x, obj.y, obj.z]).reshape(-1, 3)
bcenter_Flidar = calib.leftcam2lidar(bcenter_Fcam)
center_Flidar = bcenter_Flidar + np.array([0, 0, obj.h/2.0]).reshape(-1, 3)
if (center_Flidar[0, 0] < 0 or center_Flidar[0, 0] > 52.8
or center_Flidar[0, 1] < -30 or center_Flidar[0, 1] > 30
or center_Flidar[0, 2] < -3 or center_Flidar[0, 2] > 1):
continue
obj.truncated = 0
obj.occluded = 0
obj.alpha = 0
obj.bbox_l = 0
obj.bbox_t = 0
obj.bbox_r = 0
obj.bbox_b = 0
label_gt.add_obj(obj)
for box3d_lidar, label_preds, score in zip(
predictions_dict["box3d_lidar"],
predictions_dict["label_preds"],
predictions_dict["scores"]):
obj = KittiObj()
obj.type = "Car" if label_preds == 0 else "Pedestrian"
xyzwlhry_Flidar = box3d_lidar.cpu().numpy().flatten()
bcenter_Flidar = xyzwlhry_Flidar[:3].reshape(-1, 3)
bcenter_Fcam = calib.lidar2leftcam(bcenter_Flidar)
obj.x, obj.y, obj.z = bcenter_Fcam.flatten()
obj.w, obj.l, obj.h, obj.ry = xyzwlhry_Flidar[3:]
obj.truncated = 0
obj.occluded = 0
obj.alpha = 0
obj.bbox_l = 0
obj.bbox_t = 0
obj.bbox_r = 0
obj.bbox_b = 0
label_est.add_obj(obj)
self.assertTrue(label_gt.equal(label_est, acc_cls=["Car", "Pedestrian"], rtol=1e-2))