performance of using GT 2d box

[burui11087]

Hi.

I want to see the bottleneck of F-pointnet 3D detection module. So I replace 2d detected boundingbox with groundtruth box. But the evaluated results are strange, easy/moderate/hard 3D AP is 70.52%,65.04%,60.53% separately. Following is how I do.

I modify the command_test_v2.sh to
python3 train/test.py --gpu 0 --num_point 1024 --model frustum_pointnets_v2 --model_path train/frustum_pointnets_snapshots/log_v2/model.ckpt --output train/detection_results_v2 --data_path kitti/frustum_data/frustum_carpedcyc_val.pickle --idx_path kitti/image_sets/val.txt train/kitti_eval/evaluate_object_3d_offline dataset/KITTI/object/training/label_2/ train/detection_results_v2

I delete --from_rgb_detection and replace frustum_carpedcyc_val_rgb_detection.pickle with frustum_carpedcyc_val.pickle

Thanks!