Trained on the KITTI and test and evaluation STF

[RhythmOfTheRain-Byte]

Excuse me, I trained a model on the KITTI dataset, and now I want to test and evaluate it on another dataset called STF for domain adaptation. After loading the data, do I need to transform the STF point cloud into the KITTI coordinate system? I conducted the above experiment using SECOND-IoU. The evaluation results on the KITTI validation set were very good, but the results on the STF dataset were poor. The results are shown below, along with the evaluation-related code.

[2025-05-26 15:34:27,560 eval_utils.py 121 INFO] Car [email protected], 0.70, 0.70: bbox AP:51.9155, 51.9155, 51.9155 bev AP:0.0125, 0.0125, 0.0125 3d AP:0.0001, 0.0001, 0.0001 aos AP:27.26, 27.26, 27.26 Car [email protected], 0.70, 0.70: bbox AP:50.5135, 50.5135, 50.5135 bev AP:0.0008, 0.0008, 0.0008 3d AP:0.0000, 0.0000, 0.0000 aos AP:26.89, 26.89, 26.89 Car [email protected], 0.50, 0.50: bbox AP:51.9155, 51.9155, 51.9155 bev AP:0.1466, 0.1466, 0.1466 3d AP:0.0230, 0.0230, 0.0230 aos AP:27.26, 27.26, 27.26 Car [email protected], 0.50, 0.50: bbox AP:50.5135, 50.5135, 50.5135 bev AP:0.0360, 0.0360, 0.0360 3d AP:0.0014, 0.0014, 0.0014 aos AP:26.89, 26.89, 26.89

The evaluation-related code is as follows:

def evaluation(self, det_annos, class_names, **kwargs):
    if kwargs['eval_metric'] == 'kitti':
        eval_det_annos = copy.deepcopy(det_annos)
        # eval_gt_annos = copy.deepcopy(self.dense_infos['annos'])
        eval_gt_annos = [copy.deepcopy(info['annos']) for info in self.dense_infos]
        return self.kitti_eval(eval_det_annos, eval_gt_annos, class_names)
    # elif kwargs['eval_metric'] == 'nuscenes':
    #     return self.nuscene_eval(det_annos, class_names, **kwargs)
    else:
        raise NotImplementedError   

def kitti_eval(self, eval_det_annos, eval_gt_annos, class_names):
    from ..kitti.kitti_object_eval_python import eval as kitti_eval

    map_name_to_kitti = {
        'Pedestrian': 'Pedestrian',
        'PassengerCar': 'Car',
        'RidableVehicle': 'Cyclist',            
    } 

    def transform_to_kitti_format(annos, info_with_fakelidar=False, is_gt=False):
        for anno in annos:
            if 'name' not in anno:
                anno['name'] = anno['gt_names']
                anno.pop('gt_names')

            for k in range(anno['name'].shape[0]):
                if anno['name'][k] in map_name_to_kitti:
                    anno['name'][k] = map_name_to_kitti[anno['name'][k]]
                else:
                    anno['name'][k] = 'Person_sitting'

            if 'boxes_lidar' in anno:
                gt_boxes_lidar = anno['boxes_lidar'].copy()
            elif 'gt_boxes_lidar' in anno:
                gt_boxes_lidar = anno['gt_boxes_lidar'].copy()
            else:
                gt_boxes_lidar = anno['gt_boxes'].copy()

            # filter by fov
            if is_gt and self.dataset_cfg.get('GT_FILTER', None):
                if self.dataset_cfg.GT_FILTER.get('FOV_FILTER', None):
                    fov_gt_flag = self.extract_fov_gt(
                        gt_boxes_lidar, self.dataset_cfg['FOV_DEGREE'], self.dataset_cfg['FOV_ANGLE']
                    )
                    gt_boxes_lidar = gt_boxes_lidar[fov_gt_flag]
                    anno['name'] = anno['name'][fov_gt_flag]

            anno['bbox'] = np.zeros((len(anno['name']), 4))
            anno['bbox'][:, 2:4] = 50  # [0, 0, 50, 50]
            anno['truncated'] = np.zeros(len(anno['name']))
            anno['occluded'] = np.zeros(len(anno['name']))

            if len(gt_boxes_lidar) > 0:
                if info_with_fakelidar:
                    gt_boxes_lidar = box_utils.boxes3d_kitti_fakelidar_to_lidar(gt_boxes_lidar)

                gt_boxes_lidar[:, 2] -= gt_boxes_lidar[:, 5] / 2
                anno['location'] = np.zeros((gt_boxes_lidar.shape[0], 3))
                anno['location'][:, 0] = -gt_boxes_lidar[:, 1]  # x = -y_lidar
                anno['location'][:, 1] = -gt_boxes_lidar[:, 2]  # y = -z_lidar
                anno['location'][:, 2] = gt_boxes_lidar[:, 0]  # z = x_lidar                   
                
                dxdydz = gt_boxes_lidar[:, 3:6]
                anno['dimensions'] = dxdydz[:, [0, 2, 1]]  # lwh ==> lhw
                anno['rotation_y'] = -gt_boxes_lidar[:, 6] - np.pi / 2.0
                anno['alpha'] = -np.arctan2(-gt_boxes_lidar[:, 1], gt_boxes_lidar[:, 0]) + anno['rotation_y']
            else:
                anno['location'] = anno['dimensions'] = np.zeros((0, 3))
                anno['rotation_y'] = anno['alpha'] = np.zeros(0)

    transform_to_kitti_format(eval_det_annos)
    # transform_to_kitti_format(eval_det_annos, info_with_fakelidar=True)
    transform_to_kitti_format(eval_gt_annos, info_with_fakelidar=False, is_gt=True)
    # transform_to_kitti_format(eval_gt_annos, info_with_fakelidar=True, is_gt=True)
    

    kitti_class_names = []
    for x in class_names:
        if x in map_name_to_kitti:
            kitti_class_names.append(map_name_to_kitti[x])
        else:
            kitti_class_names.append('Person_sitting')
    ap_result_str, ap_dict = kitti_eval.get_official_eval_result(
        gt_annos=eval_gt_annos, dt_annos=eval_det_annos, current_classes=kitti_class_names
    )
    return ap_result_str, ap_dict