Gorilla-Lab-SCUT/frustum-convnet

Whats the use of 'frustum_caronly_val_rgb_detection.pickle'

Niceday99 opened this issue · 1 comments

Hi Zhixin,

What is the use of 'frustum_caronly_val_rgb_detection.pickle' dataset? It seems I can train the network without this file.

Meanwhile, how to generate 'kitti/rgb_detections/rgb_detection_val.txt', because I am using NuScenes database to train F-ConvNet, and NuScenes provide scripts to convert data structure into Kitti one, so I already generate 'train.txt' and 'val.txt' for NuScenes by myself, but I don't know how to generate 'rgb_detection_val.txt'.

We directly use the ground truth 2D results during training, but in our final evaluation, we get the final 3D detection results basing on 2D detectors. And "frustum_caronly_val_rgb_detection.pickle" is the file generating from "rgb_detection_val.txt".

The formal of "rgb_detection_val.txt" is "Image ID, Category ID, Score, Box (x1, y1, x2, y2)". You can see the source code here to understand the usage. So, for NuScene dataset, you also need a 2D detector to provide the 2D detection results and save them in above format.