关于init_proposals

[xsxyh]

根据开源代码可知，init_proposals从results_save_total.pkl获取的2D检测结果生成，可以理解为results_save_total.pkl为训练样本经过已经训练好的2D检测器获得的结果。同时在推理时，该部分逻辑也保持一致，训练样本获得的检测结果用于推理是否有作用。如果我对results_save_total.pkl的理解有误，还请告诉我results_save_total.pkl的生成方式和意义。

QAF2D/projects/mmdet3d_plugin/models/detectors/persdetr3d.py

Lines 396 to 399 in ed2b955

	if True:
	initial_proposals = [self.obtain_proposals(
	img=rec_img[:, i, ...], img_metas=img_metas[i])
	for i in range(rec_img.size(1))] # len_temp 为 len(initial_proposals)

QAF2D/projects/mmdet3d_plugin/models/detectors/persdetr3d.py

Lines 469 to 472 in ed2b955

	if True:
	initial_proposals = [self.obtain_proposals(
	img=rec_img[:, i, ...], img_metas=img_metas_temp[i], training_mode=False)
	for i in range(rec_img.size(1))] # len_temp 为 len(initial_proposals)

另一方面在推理时，并未用到2D网络，训练也时也是直接获取的results_save_total.pkl，那么该分支是否有训练的必要

QAF2D/projects/mmdet3d_plugin/models/detectors/persdetr3d.py

Lines 477 to 482 in ed2b955

	bbox_list = [dict() for i in range(len(img_metas))]
	bbox_pts = self.simple_test_pts(
	img_metas, init_proposals=initial_proposals, **data)
	for result_dict, pts_bbox in zip(bbox_list, bbox_pts):
	result_dict['pts_bbox'] = pts_bbox
	return bbox_list