Peking University/Baidu - Autonomous Driving - CenterNet

This is a fork of CenterNet repo aimed at using this architecture for aforementioned Kaggle's competition. Major changes:

added new dataset kaggle_cars;
added new pipeline car_pose_6dof;
prerendering and using 3D location masks (the idea comes from 2018 paper 3D Pose Estimation for Fine-Grained Object Categories);
mixed precision training (Nvidia's Apex amp) and gradient accumulation for higher batch size;
options to use SWA and weight decay for better generalization;

To learn more about the competition please visit kaggle.com.

Installation

Please refer to INSTALL.md for installation instructions.

(OPTIONAL) You may want to install NVIDIA Apex to be able to use mixed precision training. Please, visit this repo for more info.

Setup data

By default for --dataset kaggle_cars the script will search data in data/pku-autonomous-driving. You can create a symlink like this one:

$ ln -s /home/user/projects/kaggle_cars/input/pku-autonomous-driving /home/user/projects/kaggle_cars/centernet/data

To split images for train and validation create folder split with train.txt, val.txt, ignore.txt. Each file should contain a list of the image files with extensions. For example you may try this:

$ mkdir split
$ ls train_images | head -n 3800 > split/train.txt
$ ls train_images | tail -n 462 > split/val.txt

If you intend to use 3D location masks, you have to generate it first using another script from tools folder:

$ python src/tools/prepare_3d_loc_masks.py --num_workers 4 --norm_xyz 519.834,689.119,3502.94

By default it creates train_3d_masks folder in the dataset directory, but you can overwrite this behaviour providing another path with --xyz_masks_dir.

Run train

$ python main.py car_pose_6dof --exp_id car_pose_default --dataset kaggle_cars \
 --xyz_mask \
 --batch_size 10 --master_batch_size 5 --num_grad_accum 2 \
 --num_epochs 20 --lr 1e-4 --lr_step 15,25 --weight_decay 0.01 \
 --use_swa --swa_start 12_000 --swa_freq 20 --swa_manual \
 --aug_blur 0.25 --blur_limit 3,9 \
 --aug_noise 0.2 --noise_scale 0.03,0.09 \
 --aug_hue 0.2 --hue_shift_limit 30 \
 --aug_brightness_contrast 0.3 --brightness_limit 0.1 --contrast_limit 0.1 \
 --center_thresh 0.3 \
 --gpus 0,1

Run evaluation

Just add --test --resume flags to the command line above. If you want to load specific weights, pass it to --load_model .

Save averaged model weights

Run train script with --test --save_avg_weights. The averaged weights file name ends with _avg suffix.

Run predictions on test set

$ python test.py car_pose_6dof --exp_id car_pose_default --dataset kaggle_cars \
 --xyz_mask \
 --load_model ../exp/car_pose_6dof/car_pose_default/model_20_avg.pth --resume \
 --peak_thresh 0.5 --K 50 \
 --gpus 1 --trainval --not_prefetch_test

Debug / visualisations

Use --debug 4 --render_cars to see rendered car models. Just add these parameters to train/test command lines. Adding --debug_heatmap will also visualise heatmaps.

Citation

@inproceedings{zhou2019objects,
  title={Objects as Points},
  author={Zhou, Xingyi and Wang, Dequan and Kr{\"a}henb{\"u}hl, Philipp},
  booktitle={arXiv preprint arXiv:1904.07850},
  year={2019}
}

@inproceedings{wang20183d,
  title={3D Pose Estimation for Fine-Grained Object Categories},
  author={Wang, Yaming and Tan, Xiao and Yang, Yi and Liu, Xiao and Ding, Errui and Zhou, Feng and Davis, Larry S},
  booktitle={European Conference on Computer Vision Workshop},
  pages={619--632},
  year={2018},
  organization={Springer}
}

dmitry-vorobiev/kaggle-pku-autonomous-driving-centernet