Point GNN环境配置及运行

Point GNN配置及运行

1. 官方项目地址

GitHub - WeijingShi/Point-GNN: Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud, CVPR 2020.

---

2. 多CUDA共存

1.1 多CUDA共存

建议使用和tensorflow1.15.0版本匹配的CUDA，例如CUDA10.0，使用该版本CUDA后，可以正常运行项目，不存在报错

可参考以下链接，修改环境变量以选择使用哪个版本的CUDA

win10系统上安装多个版本的cuda - Trouvaille_fighting - 博客园 (cnblogs.com)

1.2 修改CUDA DLL名称

我的电脑装了CUDA11.1，但TensorFlow 1.x不能用到高版本的CUDA，为支持TensorFlow 1.x, 可以将CUDA一些DLL名字从11/110，改为10/100

尝试直接拷贝v11.1重命名为v10.xfake, 并修改下面的dll名称后可用
不建议使用此方法，导致了后续tensorflow session跑不了

---

3. Python3.6 VSCode 调试不可用

升级到Python3.7版本后，Python Interpreter 可使用VSCode调试

(11条消息) vscode python3.6无法debug解决方案_Johnny__Wang__的博客-CSDN博客_vscode插件降级

---

4. 因protobuf版本太高，tensorflow无法import

(11条消息) 1. Downgrade the protobuf package to 3.20.x or lower._weixin_44834086的博客-CSDN博客

---

5. 检查TensorFlow GPU支持

---

6. KITTI数据集下载

https://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d

数据集较大38.5G左右，分为几个部分，我的组织方式如下
可参考官方项目操作，左值和右值图片只用下其中一个就行

---

7. KITTI bat点云转pcd

直接读bin文件也是可以的，提供这个方法用于在cloud compare等工具查看点云
KITTI数据集之点云建图 - 灰信网（软件开发博客聚合） (freesion.com)

        int32_t num = 1000000;
        int distance_threshold=20;
	float *data = (float*)malloc(num * sizeof(float));
	// 
	float *px = data + 0;
	float *py = data + 1;
	float *pz = data + 2;
	float *pr = data + 3;//
	// 
	FILE *stream;
	fopen_s(&stream, inptfile, "rb");
	num = fread(data, sizeof(float), num, stream) / 4;//
	fclose(stream);
	int32_t after_erase_num = 0;
	int distance_threshold=20;
	float *px_t = px;
	float *py_t = py;
	float *pz_t = pz;
	float *pr_t = pr;//
	for (int32_t i = 0; i < num; i++)
	{	
		//setting a threshold according to the distance between the point and centre to decrease the points 
		double distance = sqrt((*px_t)*(*px_t) + (*py_t)*(*py_t) + (*pz_t)*(*pz_t));
		if (distance < distance_threshold) {
			fprintf(writePCDStream, "%f %f %f %f\n", *px, *py, *pz, *pr);
		}
		px_t += 4; py_t += 4; pz_t += 4; pr_t += 4;
 
	}
        fclose(writePCDStream);

---

8. 官方代码没有给3D检测框，你可以找到这部分代码，修改并查看结果

            # convert to KITTI ================================================
            detection_boxes_3d_corners = nms.boxes_3d_to_corners(
                detection_boxes_3d)
            pred_labels = []
            for i in range(len(detection_boxes_3d_corners)):
                detection_box_3d_corners = detection_boxes_3d_corners[i]
                corners_cam_points = Points(xyz=detection_box_3d_corners, attr=None)
                corners_img_points = dataset.cam_points_to_image(corners_cam_points, calib)
                corners_xy = corners_img_points.xyz[:, :2]
                if config['label_method'] == 'yaw':
                    all_class_name = ['Background', 'Car', 'Car', 'Pedestrian',
                        'Pedestrian', 'Cyclist', 'Cyclist', 'DontCare']
                if config['label_method'] == 'Car':
                    all_class_name = ['Background', 'Car', 'Car', 'DontCare']
                if config['label_method'] == 'Pedestrian_and_Cyclist':
                    all_class_name = ['Background', 'Pedestrian', 'Pedestrian',
                        'Cyclist', 'Cyclist', 'DontCare']
                if config['label_method'] == 'alpha':
                    all_class_name = ['Background', 'Car', 'Car', 'Pedestrian',
                        'Pedestrian', 'Cyclist', 'Cyclist', 'DontCare']
                class_name = all_class_name[class_labels[i]]
                # for i in range(len(corners_xy)):
                #     cv2.circle(image, (int(corners_xy[i][0]), int(corners_xy[i][1])), 3, (255, 255, 125), -1)
                # for i in range(len(corners_xy)):
                #     for j in range(len(corners_xy)):
                #         cv2.line(image, (int(corners_xy[i][0]), int(corners_xy[i][1])), (int(corners_xy[j][0]), int(corners_xy[j][1])), (0, 255, 0), 1)
                cv2.line(image, (int(corners_xy[0][0]), int(corners_xy[0][1])), (int(corners_xy[1][0]), int(corners_xy[1][1])), (0, 255, 0), 1)
                cv2.line(image, (int(corners_xy[0][0]), int(corners_xy[0][1])), (int(corners_xy[3][0]), int(corners_xy[3][1])), (0, 255, 0), 1)
                cv2.line(image, (int(corners_xy[1][0]), int(corners_xy[1][1])), (int(corners_xy[2][0]), int(corners_xy[2][1])), (0, 255, 0), 1)
                cv2.line(image, (int(corners_xy[2][0]), int(corners_xy[2][1])), (int(corners_xy[3][0]), int(corners_xy[3][1])), (0, 255, 0), 1)
                cv2.line(image, (int(corners_xy[4][0]), int(corners_xy[4][1])), (int(corners_xy[5][0]), int(corners_xy[5][1])), (0, 255, 0), 1)
                cv2.line(image, (int(corners_xy[6][0]), int(corners_xy[6][1])), (int(corners_xy[7][0]), int(corners_xy[7][1])), (0, 255, 0), 1)
                cv2.line(image, (int(corners_xy[4][0]), int(corners_xy[4][1])), (int(corners_xy[7][0]), int(corners_xy[7][1])), (0, 255, 0), 1)
                cv2.line(image, (int(corners_xy[5][0]), int(corners_xy[5][1])), (int(corners_xy[6][0]), int(corners_xy[6][1])), (0, 255, 0), 1)
                cv2.line(image, (int(corners_xy[0][0]), int(corners_xy[0][1])), (int(corners_xy[4][0]), int(corners_xy[4][1])), (0, 255, 0), 1)
                cv2.line(image, (int(corners_xy[1][0]), int(corners_xy[1][1])), (int(corners_xy[5][0]), int(corners_xy[5][1])), (0, 255, 0), 1)
                cv2.line(image, (int(corners_xy[3][0]), int(corners_xy[3][1])), (int(corners_xy[7][0]), int(corners_xy[7][1])), (0, 255, 0), 1)
                cv2.line(image, (int(corners_xy[2][0]), int(corners_xy[2][1])), (int(corners_xy[6][0]), int(corners_xy[6][1])), (0, 255, 0), 1)
                '''
                xmin, ymin = np.amin(corners_xy, axis=0)
                xmax, ymax = np.amax(corners_xy, axis=0)
                print(xmin, ymin, xmax, ymax)
                clip_xmin = max(xmin, 0.0)
                clip_ymin = max(ymin, 0.0)
                clip_xmax = min(xmax, 1242.0)
                clip_ymax = min(ymax, 375.0)
                height = clip_ymax - clip_ymin
                truncation_rate = 1.0 - (clip_ymax - clip_ymin)*(
                    clip_xmax - clip_xmin)/((ymax - ymin)*(xmax - xmin))
                if truncation_rate > 0.4:
                    continue
                x3d, y3d, z3d, l, h, w, yaw = detection_boxes_3d[i]
                assert l > 0, str(i)
                score = box_probs[i]
                if USE_BOX_SCORE:
                    tmp_label = {"x3d": x3d, "y3d" : y3d, "z3d": z3d,
                    "yaw": yaw, "height": h, "width": w, "length": l}
                    # Rescore or not ===========================================
                    inside_mask = dataset.sel_xyz_in_box3d(tmp_label,
                        last_layer_points_xyz[box_indices])
                    points_inside = last_layer_points_xyz[
                        box_indices][inside_mask]
                    score_inside = box_probs_ori[inside_mask]
                    score = (1+occlusion(tmp_label, points_inside))*score
                pred_labels.append((class_name, -1, -1, 0,
                    clip_xmin, clip_ymin, clip_xmax, clip_ymax,
                    h, w, l, x3d, y3d, z3d, yaw, score))
                if VISUALIZATION_LEVEL > 0:
                    cv2.rectangle(image,
                        (int(clip_xmin), int(clip_ymin)),
                        (int(clip_xmax), int(clip_ymax)), (0, 255, 0), 2)
                    if class_name == "Pedestrian":
                        cv2.putText(image, '{:s} | {:.3f}'.format('P', score),
                            (int(clip_xmin), int(clip_ymin)-int(clip_xmin/10)),
                            cv2.FONT_HERSHEY_SIMPLEX, 0.4, (255, 255, 0) ,1)
                    else:
                        cv2.putText(image, '{:s} | {:.3f}'.format('C', score),
                            (int(clip_xmin), int(clip_ymin)),
                            cv2.FONT_HERSHEY_SIMPLEX, 0.4, (0, 255, 0) ,1)
            '''
            nms_time = time.time()
            time_dict['nms'] = time_dict.get('nms', 0) + nms_time - decode_time

---

推理的结果