训练验证数据生成&＃xff1a;
python create_data.py nuscenes_data_prep --data_path&＃61;NUSCENES_TRAINVAL_DATASET_ROOT --version&＃61;"v1.0-trainval" --max_sweeps&＃61;10python create_data.py nuscenes_data_prep --data_path&＃61;NUSCENES_TEST_DATASET_ROOT --version&＃61;"v1.0-test" --max_sweeps&＃61;10 --dataset_name&＃61;"NuscenesDataset"
如果为自定义数据要根据类别进行修改&＃xff0c;将数据转成bin文件&＃xff0c;打包成pkl文件再执行上面指令。
如&＃xff1a; &＃xff08;1&＃xff09;my_common.py 将标签与点云打包成pkl。
&＃xff08;2&＃xff09;python my_create_data.py mydata --data_path&＃61;datapath
打开second.pytorch/second/configs/car.lite.config 编辑类别与训练数据路径
train_input_reader: {...database_sampler {database_info_path: "/path/to/dataset_dbinfos_train.pkl"...}dataset: {dataset_class_name: "DATASET_NAME"kitti_info_path: "/path/to/dataset_infos_train.pkl"kitti_root_path: "DATASET_ROOT"}
}
...
eval_input_reader: {...dataset: {dataset_class_name: "DATASET_NAME"kitti_info_path: "/path/to/dataset_infos_val.pkl"kitti_root_path: "DATASET_ROOT"}
}
单GPU&＃xff1a;
python ./pytorch/train.py train --config_path&＃61;./configs/car.fhd.config --model_dir&＃61;/path/to/model_dir
多GPU&＃xff1a;
CUDA_VISIBLE_DEVICES&＃61;0,1,3 python ./pytorch/train.py train --config_path&＃61;./configs/car.fhd.config --model_dir&＃61;/path/to/model_dir --multi_gpu&＃61;True
浮点16位训练&＃xff1a;
修改配置文件&＃xff0c;将enable\u mixed\u precision设置为true。
&＃xff08;1&＃xff09;如果要训练新模型&＃xff0c;请确保“/path/to/model\u dir”不存在。如果model\u dir不存在&＃xff0c;将创建一个新目录&＃xff0c;否则将读取其中的检查点。
&＃xff08;2&＃xff09;训练过程使用batchsize&＃61;6作为1080Ti的默认值&＃xff0c;如果GPU内存较少&＃xff0c;则需要减少batchsize。
&＃xff08;3&＃xff09;目前仅支持单个GPU训练&＃xff0c;但训练一个模型在单个1080Ti中只需要20小时&＃xff0c;并且只需要50个循环就可以达到78.3 AP&＃xff0c;并在Kitti验证日期集中使用super converge in car Medium 3D。
检测结果默认保存为&＃xff1a;result.pkl&＃xff0c;可设置--pickle_result&＃61;False&＃xff0c;将结果保存为kitt标签格式
python ./pytorch/train.py evaluate --config_path&＃61;./configs/car.fhd.config --model_dir&＃61;/path/to/model_dir --measure_time&＃61;True --batch_size&＃61;1
需要修改或者重写second.data.kitti_dataset以 &＃64;register_dataset 方式注册函数
训练时要修改eval.py&＃xff0c;主要是自定义数据的类别等。
from pathlib import Path
import pickle
import time
from functools import partialimport numpy as npfrom second.core import box_np_ops
from second.core import preprocess as prep
from second.data import kitti_common as kitti
from second.utils.eval import get_coco_eval_result, get_official_eval_result
from second.data.dataset import Dataset, register_dataset
from second.utils.progress_bar import progress_bar_iter as prog_bar&＃64;register_dataset
class KittiDataset(Dataset):NumPointFeatures &＃61; 4def __init__(self,root_path,info_path,class_names&＃61;None,prep_func&＃61;None,num_point_features&＃61;None):assert info_path is not Nonewith open(info_path, &＃39;rb&＃39;) as f:infos &＃61; pickle.load(f)self._root_path &＃61; Path(root_path)self._kitti_infos &＃61; infosprint("remain number of infos:", len(self._kitti_infos))self._class_names &＃61; class_namesself._prep_func &＃61; prep_funcdef __len__(self):return len(self._kitti_infos)def convert_detection_to_kitti_annos(self, detection):class_names &＃61; self._class_namesdet_image_idxes &＃61; [det["metadata"]["image_idx"] for det in detection]gt_image_idxes &＃61; [info["image"]["image_idx"] for info in self._kitti_infos]annos &＃61; []for i in range(len(detection)):det_idx &＃61; det_image_idxes[i]det &＃61; detection[i]# info &＃61; self._kitti_infos[gt_image_idxes.index(det_idx)]info &＃61; self._kitti_infos[i]calib &＃61; info["calib"]rect &＃61; calib["R0_rect"]Trv2c &＃61; calib["Tr_velo_to_cam"]P2 &＃61; calib["P2"]final_box_preds &＃61; det["box3d_lidar"].detach().cpu().numpy()label_preds &＃61; det["label_preds"].detach().cpu().numpy()scores &＃61; det["scores"].detach().cpu().numpy()if final_box_preds.shape[0] !&＃61; 0:final_box_preds[:, 2] -&＃61; final_box_preds[:, 5] / 2box3d_camera &＃61; box_np_ops.box_lidar_to_camera(final_box_preds, rect, Trv2c)locs &＃61; box3d_camera[:, :3]dims &＃61; box3d_camera[:, 3:6]angles &＃61; box3d_camera[:, 6]camera_box_origin &＃61; [0.5, 1.0, 0.5]box_corners &＃61; box_np_ops.center_to_corner_box3d(locs, dims, angles, camera_box_origin, axis&＃61;1)box_corners_in_image &＃61; box_np_ops.project_to_image(box_corners, P2)# box_corners_in_image: [N, 8, 2]minxy &＃61; np.min(box_corners_in_image, axis&＃61;1)maxxy &＃61; np.max(box_corners_in_image, axis&＃61;1)bbox &＃61; np.concatenate([minxy, maxxy], axis&＃61;1)anno &＃61; kitti.get_start_result_anno()num_example &＃61; 0box3d_lidar &＃61; final_box_predsfor j in range(box3d_lidar.shape[0]):image_shape &＃61; info["image"]["image_shape"]if bbox[j, 0] > image_shape[1] or bbox[j, 1] > image_shape[0]:continueif bbox[j, 2] <0 or bbox[j, 3] <0:continuebbox[j, 2:] &＃61; np.minimum(bbox[j, 2:], image_shape[::-1])bbox[j, :2] &＃61; np.maximum(bbox[j, :2], [0, 0])anno["bbox"].append(bbox[j])# convert center format to kitti format# box3d_lidar[j, 2] -&＃61; box3d_lidar[j, 5] / 2anno["alpha"].append(-np.arctan2(-box3d_lidar[j, 1], box3d_lidar[j, 0]) &＃43;box3d_camera[j, 6])anno["dimensions"].append(box3d_camera[j, 3:6])anno["location"].append(box3d_camera[j, :3])anno["rotation_y"].append(box3d_camera[j, 6])anno["name"].append(class_names[int(label_preds[j])])anno["truncated"].append(0.0)anno["occluded"].append(0)anno["score"].append(scores[j])num_example &＃43;&＃61; 1if num_example !&＃61; 0:anno &＃61; {n: np.stack(v) for n, v in anno.items()}annos.append(anno)else:annos.append(kitti.empty_result_anno())num_example &＃61; annos[-1]["name"].shape[0]annos[-1]["metadata"] &＃61; det["metadata"]return annosdef evaluation(self, detections, output_dir):"""detectionWhen you want to eval your own dataset, you MUST set correctthe z axis and box z center.If you want to eval by my KITTI eval function, you must provide the correct format annotations.ground_truth_annotations format:{bbox: [N, 4], if you fill fake data, MUST HAVE >25 HEIGHT!!!!!!alpha: [N], you can use -10 to ignore it.occluded: [N], you can use zero.truncated: [N], you can use zero.name: [N]location: [N, 3] center of 3d box.dimensions: [N, 3] dim of 3d box.rotation_y: [N] angle.}all fields must be filled, but some fields can fillzero."""if "annos" not in self._kitti_infos[0]:return Nonegt_annos &＃61; [info["annos"] for info in self._kitti_infos]dt_annos &＃61; self.convert_detection_to_kitti_annos(detections)# firstly convert standard detection to kitti-format dt annosz_axis &＃61; 1 # KITTI camera format use y as regular "z" axis.z_center &＃61; 1.0 # KITTI camera box&＃39;s center is [0.5, 1, 0.5]# for regular raw lidar data, z_axis &＃61; 2, z_center &＃61; 0.5.result_official_dict &＃61; get_official_eval_result(gt_annos,dt_annos,self._class_names,z_axis&＃61;z_axis,z_center&＃61;z_center)result_coco &＃61; get_coco_eval_result(gt_annos,dt_annos,self._class_names,z_axis&＃61;z_axis,z_center&＃61;z_center)return {"results": {"official": result_official_dict["result"],"coco": result_coco["result"],},"detail": {"eval.kitti": {"official": result_official_dict["detail"],"coco": result_coco["detail"]}},}def __getitem__(self, idx):input_dict &＃61; self.get_sensor_data(idx)example &＃61; self._prep_func(input_dict&＃61;input_dict)example["metadata"] &＃61; {}if "image_idx" in input_dict["metadata"]:example["metadata"] &＃61; input_dict["metadata"]if "anchors_mask" in example:example["anchors_mask"] &＃61; example["anchors_mask"].astype(np.uint8)return exampledef get_sensor_data(self, query):read_image &＃61; Falseidx &＃61; queryif isinstance(query, dict):read_image &＃61; "cam" in queryassert "lidar" in queryidx &＃61; query["lidar"]["idx"]info &＃61; self._kitti_infos[idx]res &＃61; {"lidar": {"type": "lidar","points": None,},"metadata": {"image_idx": info["image"]["image_idx"],"image_shape": info["image"]["image_shape"],},"calib": None,"cam": {}}pc_info &＃61; info["point_cloud"]velo_path &＃61; Path(pc_info[&＃39;velodyne_path&＃39;])if not velo_path.is_absolute():velo_path &＃61; Path(self._root_path) / pc_info[&＃39;velodyne_path&＃39;]velo_reduced_path &＃61; velo_path.parent.parent / (velo_path.parent.stem &＃43; &＃39;_reduced&＃39;) / velo_path.nameif velo_reduced_path.exists():velo_path &＃61; velo_reduced_pathpoints &＃61; np.fromfile(str(velo_path), dtype&＃61;np.float32,count&＃61;-1).reshape([-1, self.NumPointFeatures])res["lidar"]["points"] &＃61; pointsimage_info &＃61; info["image"]image_path &＃61; image_info[&＃39;image_path&＃39;]if read_image:image_path &＃61; self._root_path / image_pathwith open(str(image_path), &＃39;rb&＃39;) as f:image_str &＃61; f.read()res["cam"] &＃61; {"type": "camera","data": image_str,"datatype": image_path.suffix[1:],}calib &＃61; info["calib"]calib_dict &＃61; {&＃39;rect&＃39;: calib[&＃39;R0_rect&＃39;],&＃39;Trv2c&＃39;: calib[&＃39;Tr_velo_to_cam&＃39;],&＃39;P2&＃39;: calib[&＃39;P2&＃39;],}res["calib"] &＃61; calib_dictif &＃39;annos&＃39; in info:annos &＃61; info[&＃39;annos&＃39;]# we need other objects to avoid collision when sampleannos &＃61; kitti.remove_dontcare(annos)locs &＃61; annos["location"]dims &＃61; annos["dimensions"]rots &＃61; annos["rotation_y"]gt_names &＃61; annos["name"]# rots &＃61; np.concatenate([np.zeros([locs.shape[0], 2], dtype&＃61;np.float32), rots], axis&＃61;1)gt_boxes &＃61; np.concatenate([locs, dims, rots[..., np.newaxis]],axis&＃61;1).astype(np.float32)calib &＃61; info["calib"]gt_boxes &＃61; box_np_ops.box_camera_to_lidar(gt_boxes, calib["R0_rect"], calib["Tr_velo_to_cam"])# only center format is allowed. so we need to convert# kitti [0.5, 0.5, 0] center to [0.5, 0.5, 0.5]box_np_ops.change_box3d_center_(gt_boxes, [0.5, 0.5, 0],[0.5, 0.5, 0.5])res["lidar"]["annotations"] &＃61; {&＃39;boxes&＃39;: gt_boxes,&＃39;names&＃39;: gt_names,}res["cam"]["annotations"] &＃61; {&＃39;boxes&＃39;: annos["bbox"],&＃39;names&＃39;: gt_names,}return resdef convert_to_kitti_info_version2(info):"""convert kitti info v1 to v2 if possible."""if "image" not in info or "calib" not in info or "point_cloud" not in info:info["image"] &＃61; {&＃39;image_shape&＃39;: info["img_shape"],&＃39;image_idx&＃39;: info[&＃39;image_idx&＃39;],&＃39;image_path&＃39;: info[&＃39;img_path&＃39;],}info["calib"] &＃61; {"R0_rect": info[&＃39;calib/R0_rect&＃39;],"Tr_velo_to_cam": info[&＃39;calib/Tr_velo_to_cam&＃39;],"P2": info[&＃39;calib/P2&＃39;],}info["point_cloud"] &＃61; {"velodyne_path": info[&＃39;velodyne_path&＃39;],}def kitti_anno_to_label_file(annos, folder):folder &＃61; Path(folder)for anno in annos:image_idx &＃61; anno["metadata"]["image_idx"]label_lines &＃61; []for j in range(anno["bbox"].shape[0]):label_dict &＃61; {&＃39;name&＃39;: anno["name"][j],&＃39;alpha&＃39;: anno["alpha"][j],&＃39;bbox&＃39;: anno["bbox"][j],&＃39;location&＃39;: anno["location"][j],&＃39;dimensions&＃39;: anno["dimensions"][j],&＃39;rotation_y&＃39;: anno["rotation_y"][j],&＃39;score&＃39;: anno["score"][j],}label_line &＃61; kitti.kitti_result_line(label_dict)label_lines.append(label_line)label_file &＃61; folder / f"{kitti.get_image_index_str(image_idx)}.txt"label_str &＃61; &＃39;\n&＃39;.join(label_lines)with open(label_file, &＃39;w&＃39;) as f:f.write(label_str)def _read_imageset_file(path):with open(path, &＃39;r&＃39;) as f:lines &＃61; f.readlines()return [int(line) for line in lines]def _calculate_num_points_in_gt(data_path,infos,relative_path,remove_outside&＃61;True,num_features&＃61;4):for info in infos:pc_info &＃61; info["point_cloud"]image_info &＃61; info["image"]calib &＃61; info["calib"]if relative_path:v_path &＃61; str(Path(data_path) / pc_info["velodyne_path"])else:v_path &＃61; pc_info["velodyne_path"]points_v &＃61; np.fromfile(v_path, dtype&＃61;np.float32, count&＃61;-1).reshape([-1, num_features])rect &＃61; calib[&＃39;R0_rect&＃39;]Trv2c &＃61; calib[&＃39;Tr_velo_to_cam&＃39;]P2 &＃61; calib[&＃39;P2&＃39;]if remove_outside:points_v &＃61; box_np_ops.remove_outside_points(points_v, rect, Trv2c, P2, image_info["image_shape"])# points_v &＃61; points_v[points_v[:, 0] > 0]annos &＃61; info[&＃39;annos&＃39;]num_obj &＃61; len([n for n in annos[&＃39;name&＃39;] if n !&＃61; &＃39;DontCare&＃39;])# annos &＃61; kitti.filter_kitti_anno(annos, [&＃39;DontCare&＃39;])dims &＃61; annos[&＃39;dimensions&＃39;][:num_obj]loc &＃61; annos[&＃39;location&＃39;][:num_obj]rots &＃61; annos[&＃39;rotation_y&＃39;][:num_obj]gt_boxes_camera &＃61; np.concatenate([loc, dims, rots[..., np.newaxis]],axis&＃61;1)gt_boxes_lidar &＃61; box_np_ops.box_camera_to_lidar(gt_boxes_camera, rect, Trv2c)indices &＃61; box_np_ops.points_in_rbbox(points_v[:, :3], gt_boxes_lidar)num_points_in_gt &＃61; indices.sum(0)num_ignored &＃61; len(annos[&＃39;dimensions&＃39;]) - num_objnum_points_in_gt &＃61; np.concatenate([num_points_in_gt, -np.ones([num_ignored])])annos["num_points_in_gt"] &＃61; num_points_in_gt.astype(np.int32)def create_kitti_info_file(data_path, save_path&＃61;None, relative_path&＃61;True):imageset_folder &＃61; Path(__file__).resolve().parent / "ImageSets"train_img_ids &＃61; _read_imageset_file(str(imageset_folder / "train.txt"))val_img_ids &＃61; _read_imageset_file(str(imageset_folder / "val.txt"))test_img_ids &＃61; _read_imageset_file(str(imageset_folder / "test.txt"))print("Generate info. this may take several minutes.")if save_path is None:save_path &＃61; Path(data_path)else:save_path &＃61; Path(save_path)kitti_infos_train &＃61; kitti.get_kitti_image_info(data_path,training&＃61;True,velodyne&＃61;True,calib&＃61;True,image_ids&＃61;train_img_ids,relative_path&＃61;relative_path)_calculate_num_points_in_gt(data_path, kitti_infos_train, relative_path)filename &＃61; save_path / &＃39;kitti_infos_train.pkl&＃39;print(f"Kitti info train file is saved to {filename}")with open(filename, &＃39;wb&＃39;) as f:pickle.dump(kitti_infos_train, f)kitti_infos_val &＃61; kitti.get_kitti_image_info(data_path,training&＃61;True,velodyne&＃61;True,calib&＃61;True,image_ids&＃61;val_img_ids,relative_path&＃61;relative_path)_calculate_num_points_in_gt(data_path, kitti_infos_val, relative_path)filename &＃61; save_path / &＃39;kitti_infos_val.pkl&＃39;print(f"Kitti info val file is saved to {filename}")with open(filename, &＃39;wb&＃39;) as f:pickle.dump(kitti_infos_val, f)filename &＃61; save_path / &＃39;kitti_infos_trainval.pkl&＃39;print(f"Kitti info trainval file is saved to {filename}")with open(filename, &＃39;wb&＃39;) as f:pickle.dump(kitti_infos_train &＃43; kitti_infos_val, f)kitti_infos_test &＃61; kitti.get_kitti_image_info(data_path,training&＃61;False,label_info&＃61;False,velodyne&＃61;True,calib&＃61;True,image_ids&＃61;test_img_ids,relative_path&＃61;relative_path)filename &＃61; save_path / &＃39;kitti_infos_test.pkl&＃39;print(f"Kitti info test file is saved to {filename}")with open(filename, &＃39;wb&＃39;) as f:pickle.dump(kitti_infos_test, f)def _create_reduced_point_cloud(data_path,info_path,save_path&＃61;None,back&＃61;False):with open(info_path, &＃39;rb&＃39;) as f:kitti_infos &＃61; pickle.load(f)for info in prog_bar(kitti_infos):pc_info &＃61; info["point_cloud"]image_info &＃61; info["image"]calib &＃61; info["calib"]v_path &＃61; pc_info[&＃39;velodyne_path&＃39;]v_path &＃61; Path(data_path) / v_pathpoints_v &＃61; np.fromfile(str(v_path), dtype&＃61;np.float32, count&＃61;-1).reshape([-1, 4])rect &＃61; calib[&＃39;R0_rect&＃39;]P2 &＃61; calib[&＃39;P2&＃39;]Trv2c &＃61; calib[&＃39;Tr_velo_to_cam&＃39;]# first remove z <0 points# keep &＃61; points_v[:, -1] > 0# points_v &＃61; points_v[keep]# then remove outside.if back:points_v[:, 0] &＃61; -points_v[:, 0]points_v &＃61; box_np_ops.remove_outside_points(points_v, rect, Trv2c, P2,image_info["image_shape"])if save_path is None:save_filename &＃61; v_path.parent.parent / (v_path.parent.stem &＃43; "_reduced") / v_path.name# save_filename &＃61; str(v_path) &＃43; &＃39;_reduced&＃39;if back:save_filename &＃43;&＃61; "_back"else:save_filename &＃61; str(Path(save_path) / v_path.name)if back:save_filename &＃43;&＃61; "_back"with open(save_filename, &＃39;w&＃39;) as f:points_v.tofile(f)def create_reduced_point_cloud(data_path,train_info_path&＃61;None,val_info_path&＃61;None,test_info_path&＃61;None,save_path&＃61;None,with_back&＃61;False):if train_info_path is None:train_info_path &＃61; Path(data_path) / &＃39;kitti_infos_train.pkl&＃39;if val_info_path is None:val_info_path &＃61; Path(data_path) / &＃39;kitti_infos_val.pkl&＃39;if test_info_path is None:test_info_path &＃61; Path(data_path) / &＃39;kitti_infos_test.pkl&＃39;_create_reduced_point_cloud(data_path, train_info_path, save_path)_create_reduced_point_cloud(data_path, val_info_path, save_path)_create_reduced_point_cloud(data_path, test_info_path, save_path)if with_back:_create_reduced_point_cloud(data_path, train_info_path, save_path, back&＃61;True)_create_reduced_point_cloud(data_path, val_info_path, save_path, back&＃61;True)_create_reduced_point_cloud(data_path, test_info_path, save_path, back&＃61;True)if __name__ &＃61;&＃61; "__main__":fire.Fire()
修改second.utils.eval class_to_name &＃61; {
0: &＃39;Car&＃39;,
1: &＃39;Pedestrian&＃39;,
2: &＃39;Cyclist&＃39;,
3: &＃39;Van&＃39;,
4: &＃39;Person_sitting&＃39;,
5: &＃39;car&＃39;,
6: &＃39;tractor&＃39;,
7: &＃39;trailer&＃39;,
}
def get_official_eval_result(gt_annos,dt_annos,current_classes,difficultys&＃61;[0, 1, 2],z_axis&＃61;1,z_center&＃61;1.0):"""gt_annos and dt_annos must contains following keys:[bbox, location, dimensions, rotation_y, score]"""overlap_mod &＃61; np.array([[0.7, 0.5, 0.5, 0.7, 0.5, 0.7, 0.7, 0.7],[0.7, 0.5, 0.5, 0.7, 0.5, 0.7, 0.7, 0.7],[0.7, 0.5, 0.5, 0.7, 0.5, 0.7, 0.7, 0.7]])overlap_easy &＃61; np.array([[0.7, 0.5, 0.5, 0.7, 0.5, 0.5, 0.5, 0.5],[0.5, 0.25, 0.25, 0.5, 0.25, 0.5, 0.5, 0.5],[0.5, 0.25, 0.25, 0.5, 0.25, 0.5, 0.5, 0.5]])min_overlaps &＃61; np.stack([overlap_mod, overlap_easy], axis&＃61;0) # [2, 3, 5]class_to_name &＃61; {0: &＃39;Car&＃39;,1: &＃39;Pedestrian&＃39;,2: &＃39;Cyclist&＃39;,3: &＃39;Van&＃39;,4: &＃39;Person_sitting&＃39;,5: &＃39;car&＃39;,6: &＃39;tractor&＃39;,7: &＃39;trailer&＃39;,}name_to_class &＃61; {v: n for n, v in class_to_name.items()}if not isinstance(current_classes, (list, tuple)):current_classes &＃61; [current_classes]current_classes_int &＃61; []for curcls in current_classes:if isinstance(curcls, str):current_classes_int.append(name_to_class[curcls])else:current_classes_int.append(curcls)current_classes &＃61; current_classes_intmin_overlaps &＃61; min_overlaps[:, :, current_classes]result &＃61; &＃39;&＃39;# check whether alpha is validcompute_aos &＃61; Falsefor anno in dt_annos:if anno[&＃39;alpha&＃39;].shape[0] !&＃61; 0:if anno[&＃39;alpha&＃39;][0] !&＃61; -10:compute_aos &＃61; Truebreakmetrics &＃61; do_eval_v3(gt_annos,dt_annos,current_classes,min_overlaps,compute_aos,difficultys,z_axis&＃61;z_axis,z_center&＃61;z_center)detail &＃61; {}for j, curcls in enumerate(current_classes):# mAP threshold array: [num_minoverlap, metric, class]# mAP result: [num_class, num_diff, num_minoverlap]class_name &＃61; class_to_name[curcls]detail[class_name] &＃61; {}for i in range(min_overlaps.shape[0]):mAPbbox &＃61; get_mAP(metrics["bbox"]["precision"][j, :, i])mAPbev &＃61; get_mAP(metrics["bev"]["precision"][j, :, i])mAP3d &＃61; get_mAP(metrics["3d"]["precision"][j, :, i])detail[class_name][f"bbox&＃64;{min_overlaps[i, 0, j]:.2f}"] &＃61; mAPbbox.tolist()detail[class_name][f"bev&＃64;{min_overlaps[i, 1, j]:.2f}"] &＃61; mAPbev.tolist()detail[class_name][f"3d&＃64;{min_overlaps[i, 2, j]:.2f}"] &＃61; mAP3d.tolist()result &＃43;&＃61; print_str((f"{class_to_name[curcls]} ""AP(Average Precision)&＃64;{:.2f}, {:.2f}, {:.2f}:".format(*min_overlaps[i, :, j])))mAPbbox &＃61; ", ".join(f"{v:.2f}" for v in mAPbbox)mAPbev &＃61; ", ".join(f"{v:.2f}" for v in mAPbev)mAP3d &＃61; ", ".join(f"{v:.2f}" for v in mAP3d)result &＃43;&＃61; print_str(f"bbox AP:{mAPbbox}")result &＃43;&＃61; print_str(f"bev AP:{mAPbev}")result &＃43;&＃61; print_str(f"3d AP:{mAP3d}")if compute_aos:mAPaos &＃61; get_mAP(metrics["bbox"]["orientation"][j, :, i])detail[class_name][f"aos"] &＃61; mAPaos.tolist()mAPaos &＃61; ", ".join(f"{v:.2f}" for v in mAPaos)result &＃43;&＃61; print_str(f"aos AP:{mAPaos}")return {"result": result,"detail": detail,}
京公网安备 11010802041100号 | 京ICP备19059560号-4 | PHP1.CN 第一PHP社区 版权所有