SYNTHIA_FPNet

Frustum PointNet for 3D Object Detection from Point Clouds using SYNTHIA Dataset

Steps

Step 0. Convert SYNTHIA to KITTI on server

0.1 . Change directory structures in code/new_synthia_to_kitti.py as follows:

# Changes in data dir
# Read Operations --> change according to need

data_dir = '/../../datatmp/Datasets/detection/SynthiaReloaded/50_model_illumination/'

# Write Operations --> change according to need
  
IMG_ROOT = '/../../datatmp/Datasets/detection/SynthiaReloaded/SYNTHIA_KITTI_1/object/training/image_2/'

PC_ROOT = '/../../datatmp/Datasets/detection/SynthiaReloaded/SYNTHIA_KITTI_1/object/training/velodyne/'

CALIB_ROOT = '/../../datatmp/Datasets/detection/SynthiaReloaded/SYNTHIA_KITTI_1/object/training/calib/'

LABEL_ROOT = '/../../datatmp/Datasets/detection/SynthiaReloaded/SYNTHIA_KITTI_1/object/training/label_2/'

# Comment following if not needed
ANOT_IMG_ROOT = '/../../datatmp/Datasets/detection/SynthiaReloaded/SYNTHIA_KITTI_1/object/training/anot_image_2/'

0.2. Run following from code:

python3 new_synthia_to_kitti.py

Step 1. Generate train pickle locally

NOTE: All the following changes (for 1.) are for local/syn_fpnet

1.1. Copy data from server (SynthiaReloaded/SYNTHIA_KITTI_1) to local machine in folder syn_fpnet/dataset (and rename to KITTI)

1.2. Change (write) image_sets/train.txt to required number of training samples (script provided in SYNTHIA_KITTI_conversion/gen_traintxt.py)

1.3. Change script command_prep_data.sh as follows to generate only train pickle:

python kitti/prepare_data.py --gen_train 

1.4. Change kitti/kitti_object.py line number 34 to required number of training samples as follows:

self.num_samples = 8088

1.5. Then to prepare the data, simply run:

sh scripts/command_prep_data.sh

Step 2.

2.1 Copy val pickle files from (previous work) server/new_fpnet/kitti folder to server/syn_fpnet/kitti

Step 3. Train on server

3.1. Open screen in server 3.2. Make the following changes for GPU number to scripts/command_train_v1.sh:

python train/train.py --gpu 4 --model frustum_pointnets_v1 --log_dir train/log_v1 --num_point 1024 --max_epoch 201 --batch_size 32 --decay_step 800000 --decay_rate 0.5

3.3. Then run the following:

CUDA_VISIBLE_DEVICES=4 sh scripts/command_train_v1.sh

Visualization of Point Clouds with bounding boxes

Run the following from syn_fpnet:

python kitti/prepare_data.py --demo