SECOND for KITTI/NuScenes object detection

ADD TANET attention module for pointpillars

(ref paper: TANet: Robust 3D Object Detection from Point Clouds with Triple Attention

ref repo: https://github.com/happinesslz/TANet)

ADD Multiclass-balanced module

(ref repo: https://github.com/poodarchu/Det3D)

ADD Lyft dataset info generate module(may exist some bugs in it)

there still exist some little fix in the pointpillars feature extractor for more convenience to depoly with TensorRT.

Have tried to deploy pointpillars with ncnn(https://github.com/Tencent/ncnn)

The inference framework memory cost is very low and not using CUDA(it use vulkan)

Have convert pytorch model --> onnx --> ncnn model

Unfortunately, have no extra time to continue this work.

My tested environment:

Python3.6.8

pytorch1.0 / pytorch1.3.0 / pytorch1.3.1

CUDA 10.0 / CUDA 10.1

Ubuntu16.04

If you want to train nuscenes dataset, see this.

News

2019-4-1: SECOND V1.6.0alpha released: New Data API, NuScenes support, PointPillars support, fp16 and multi-gpu support.

2019-3-21: SECOND V1.5.1 (minor improvement and bug fix) released!

2019-1-20: SECOND V1.5 released! Sparse convolution-based network.

See release notes for more details.

WARNING: you should rerun info generation after every code update.

Performance in KITTI validation set (50/50 split)

car.fhd.config + 160 epochs (25 fps in 1080Ti):

Car AP@0.70, 0.70, 0.70:
bbox AP:90.77, 89.50, 80.80
bev  AP:90.28, 87.73, 79.67
3d   AP:88.84, 78.43, 76.88

car.fhd.config + 50 epochs + super converge (6.5 hours) + (25 fps in 1080Ti):

Car AP@0.70, 0.70, 0.70:
bbox AP:90.78, 89.59, 88.42
bev  AP:90.12, 87.87, 86.77
3d   AP:88.62, 78.31, 76.62

car.fhd.onestage.config + 50 epochs + super converge (6.5 hours) + (25 fps in 1080Ti):

Car AP@0.70, 0.70, 0.70:
bbox AP:97.65, 89.59, 88.72
bev  AP:90.38, 88.20, 86.98
3d   AP:89.16, 78.78, 77.41

Performance in NuScenes validation set (all.pp.config, NuScenes mini train set, 3517 samples, not v1.0-mini)

car Nusc dist AP@0.5, 1.0, 2.0, 4.0
62.90, 73.07, 76.77, 78.79
bicycle Nusc dist AP@0.5, 1.0, 2.0, 4.0
0.00, 0.00, 0.00, 0.00
bus Nusc dist AP@0.5, 1.0, 2.0, 4.0
9.53, 26.17, 38.01, 40.60
construction_vehicle Nusc dist AP@0.5, 1.0, 2.0, 4.0
0.00, 0.00, 0.44, 1.43
motorcycle Nusc dist AP@0.5, 1.0, 2.0, 4.0
9.25, 12.90, 13.69, 14.11
pedestrian Nusc dist AP@0.5, 1.0, 2.0, 4.0
61.44, 62.61, 64.09, 66.35
traffic_cone Nusc dist AP@0.5, 1.0, 2.0, 4.0
11.63, 13.14, 15.81, 21.22
trailer Nusc dist AP@0.5, 1.0, 2.0, 4.0
0.80, 9.90, 17.61, 23.26
truck Nusc dist AP@0.5, 1.0, 2.0, 4.0
9.81, 21.40, 27.55, 30.34

Install

1. Clone code

git clone https://github.com/traveller59/second.pytorch.git
cd ./second.pytorch/second

2. Install dependence python packages

It is recommend to use Anaconda package manager.

conda install scikit-image scipy numba pillow matplotlib

pip install fire tensorboardX protobuf opencv-python

If you don't have Anaconda:

pip install numba scikit-image scipy pillow

Follow instructions in spconv to install spconv.

If you want to train with fp16 mixed precision (train faster in RTX series, Titan V/RTX and Tesla V100, but I only have 1080Ti), you need to install apex.

If you want to use NuScenes dataset, you need to install nuscenes-devkit.

3. Setup cuda for numba (will be removed in 1.6.0 release)

you need to add following environment variable for numba.cuda, you can add them to ~/.bashrc:

export NUMBAPRO_CUDA_DRIVER=/usr/lib/x86_64-linux-gnu/libcuda.so
export NUMBAPRO_NVVM=/usr/local/cuda/nvvm/lib64/libnvvm.so
export NUMBAPRO_LIBDEVICE=/usr/local/cuda/nvvm/libdevice

4. add second.pytorch/ to PYTHONPATH

Prepare dataset

• KITTI Dataset preparation

Download KITTI dataset and create some directories first:

└── KITTI_DATASET_ROOT
       ├── training    <-- 7481 train data
       |   ├── image_2 <-- for visualization
       |   ├── calib
       |   ├── label_2
       |   ├── velodyne
       |   └── velodyne_reduced <-- empty directory
       └── testing     <-- 7580 test data
           ├── image_2 <-- for visualization
           ├── calib
           ├── velodyne
           └── velodyne_reduced <-- empty directory

Then run

python create_data.py kitti_data_prep --data_path=KITTI_DATASET_ROOT

• NuScenes Dataset preparation

Download NuScenes dataset:

└── NUSCENES_TRAINVAL_DATASET_ROOT
       ├── samples       <-- key frames
       ├── sweeps        <-- frames without annotation
       ├── maps          <-- unused
       └── v1.0-trainval <-- metadata and annotations
└── NUSCENES_TEST_DATASET_ROOT
       ├── samples       <-- key frames
       ├── sweeps        <-- frames without annotation
       ├── maps          <-- unused
       └── v1.0-test     <-- metadata

Then run

python create_data.py nuscenes_data_prep --data_path=NUSCENES_TRAINVAL_DATASET_ROOT --version="v1.0-trainval" --max_sweeps=10
python create_data.py nuscenes_data_prep --data_path=NUSCENES_TEST_DATASET_ROOT --version="v1.0-test" --max_sweeps=10
--dataset_name="NuscenesDataset"

This will create gt database without velocity. to add velocity, use dataset name NuscenesDatasetVelo.

• Modify config file

There is some path need to be configured in config file:

train_input_reader: {
  ...
  database_sampler {
    database_info_path: "/path/to/dataset_dbinfos_train.pkl"
    ...
  }
  dataset: {
    dataset_class_name: "DATASET_NAME"
    kitti_info_path: "/path/to/dataset_infos_train.pkl"
    kitti_root_path: "DATASET_ROOT"
  }
}
...
eval_input_reader: {
  ...
  dataset: {
    dataset_class_name: "DATASET_NAME"
    kitti_info_path: "/path/to/dataset_infos_val.pkl"
    kitti_root_path: "DATASET_ROOT"
  }
}

Usage

train

I recommend to use script.py to train and eval. see script.py for more details.

train with single GPU

python ./pytorch/train.py train --config_path=./configs/car.fhd.config --model_dir=/path/to/model_dir

train with multiple GPU (need test, I only have one GPU)

Assume you have 4 GPUs and want to train with 3 GPUs:

CUDA_VISIBLE_DEVICES=0,1,3 python ./pytorch/train.py train --config_path=./configs/car.fhd.config --model_dir=/path/to/model_dir --multi_gpu=True

Note: The batch_size and num_workers in config file is per-GPU, if you use multi-gpu, they will be multiplied by number of GPUs. Don't modify them manually.

You need to modify total step in config file. For example, 50 epochs = 15500 steps for car.lite.config and single GPU, if you use 4 GPUs, you need to divide steps and steps_per_eval by 4.

train with fp16 (mixed precision)

Modify config file, set enable_mixed_precision to true.

• Make sure "/path/to/model_dir" doesn't exist if you want to train new model. A new directory will be created if the model_dir doesn't exist, otherwise will read checkpoints in it.

• training process use batchsize=6 as default for 1080Ti, you need to reduce batchsize if your GPU has less memory.

• Currently only support single GPU training, but train a model only needs 20 hours (165 epoch) in a single 1080Ti and only needs 50 epoch to reach 78.3 AP with super converge in car moderate 3D in Kitti validation dateset.

evaluate

python ./pytorch/train.py evaluate --config_path=./configs/car.fhd.config --model_dir=/path/to/model_dir --measure_time=True --batch_size=1

• detection result will saved as a result.pkl file in model_dir/eval_results/step_xxx or save as official KITTI label format if you use --pickle_result=False.

pretrained model

You can download pretrained models in google drive. The car_fhd model is corresponding to car.fhd.config.

Note that this pretrained model is trained before a bug of sparse convolution fixed, so the eval result may slightly worse.

Docker (Deprecated. I can't push docker due to network problem.)

You can use a prebuilt docker for testing:

docker pull scrin/second-pytorch 

Then run:

nvidia-docker run -it --rm -v /media/yy/960evo/datasets/:/root/data -v $HOME/pretrained_models:/root/model --ipc=host second-pytorch:latest
python ./pytorch/train.py evaluate --config_path=./configs/car.config --model_dir=/root/model/car

Try Kitti Viewer Web

Major step

1. run python ./kittiviewer/backend/main.py main --port=xxxx in your server/local.

2. run cd ./kittiviewer/frontend && python -m http.server to launch a local web server.

3. open your browser and enter your frontend url (e.g. http://127.0.0.1:8000, default]).

4. input backend url (e.g. http://127.0.0.1:16666)

5. input root path, info path and det path (optional)

6. click load, loadDet (optional), input image index in center bottom of screen and press Enter.

Inference step

Firstly the load button must be clicked and load successfully.

1. input checkpointPath and configPath.

2. click buildNet.

3. click inference.

Concepts

• Kitti lidar box

A kitti lidar box is consist of 7 elements: [x, y, z, w, l, h, rz], see figure.

All training and inference code use kitti box format. So we need to convert other format to KITTI format before training.

• Kitti camera box

A kitti camera box is consist of 7 elements: [x, y, z, l, h, w, ry].