Skip to content

E2EC: An End-to-End Contour-based Method for High-Quality High-Speed Instance Segmentation

License

Notifications You must be signed in to change notification settings

zhang-tao-whu/e2ec

Repository files navigation

E2EC: An End-to-End Contour-based Method for High-Quality High-Speed Instance Segmentation

city

E2EC: An End-to-End Contour-based Method for High-Quality High-Speed Instance Segmentation
Tao Zhang, Shiqing Wei, Shunping Ji
CVPR 2022

Any questions or discussions are welcomed!

News

An e2ec implementation using mmdetection is complete: e2ec-mmdet. The performance on coco-val is as follows:

Detector Agu shcedule AP config
FCOS w/o 32.9 config
FCOS w 35.1 config

Installation

Please see INSTALL.md.

Performances

We re-tested the speed on a single RTX3090.

Dtataset AP Image size FPS
SBD val 59.2 512×512 59.60
COCO test-dev 33.8 original size 35.25
KINS val 34.0 768×2496 12.39
Cityscapes val 34.0 1216×2432 8.58

The accuracy and inference speed of the contours at different stages on SBD val set. We also re-tested the speed on a single RTX3090.

stage init coarse final final-dml
AP 51.4 55.9 58.8 59.2
FPS 101.73 91.35 67.48 59.6

The accuracy and inference speed of the contours at different stages on coco val set.

stage init coarse final final-dml
AP 27.8 31.6 33.5 33.6
FPS 80.97 72.81 42.55 35.25

Testing

Testing on COCO

  1. Download the pretrained model here or Baiduyun(password is e2ec).

  2. Prepared the COCO dataset according to the INSTALL.md.

  3. Test:

    # testing segmentation accuracy on coco val set
    python test.py coco --checkpoint /path/to/model_coco.pth --with_nms True
    # testing detection accuracy on coco val set
    python test.py coco --checkpoint /path/to/model_coco.pth --with_nms True --eval bbox
    # testing the speed
    python test.py coco --checkpoint /path/to/model_coco.pth --with_nms True --type speed
    # testing the contours of specified stage(init/coarse/final/final-dml)
    python test.py coco --checkpoint /path/to/model_coco.pth --with_nms True --stage coarse
    # testing on coco test-dev set, run and submit data/result/results.json
    python test.py coco --checkpoint /path/to/model_coco.pth --with_nms True --dataset coco_test
    

Testing on SBD

  1. Download the pretrained model here or Baiduyun(password is e2ec).

  2. Prepared the SBD dataset according to the INSTALL.md.

  3. Test:

    # testing segmentation accuracy on SBD
    python test.py sbd --checkpoint /path/to/model_sbd.pth
    # testing detection accuracy on SBD
    python test.py sbd --checkpoint /path/to/model_sbd.pth --eval bbox
    # testing the speed
    python test.py sbd --checkpoint /path/to/model_sbd.pth --type speed
    # testing the contours of specified stage(init/coarse/final/final-dml)
    python test.py sbd --checkpoint /path/to/model_sbd.pth --stage coarse
    

Testing on KINS

  1. Download the pretrained model here or Baiduyun(password is e2ec).

  2. Prepared the KINS dataset according to the INSTALL.md.

  3. Test:

    Maybe you will find some troules, such as object of type <class 'numpy.float64'> cannot be safely interpreted as an integer. Please modify the /path/to/site-packages/pycocotools/cooceval.py. Replace np.round((0.95 - .5) / .05) in lines 506 and 507 with int(np.round((0.95 - .5) / .05)).

    # testing segmentation accuracy on KINS
    python test.py kitti --checkpoint /path/to/model_kitti.pth
    # testing detection accuracy on KINS
    python test.py kitti --checkpoint /path/to/model_kitti.pth --eval bbox
    # testing the speed
    python test.py kitti --checkpoint /path/to/model_kitti.pth --type speed
    # testing the contours of specified stage(init/coarse/final/final-dml)
    python test.py kitti --checkpoint /path/to/model_kitti.pth --stage coarse
    

Testing on Cityscapes

  1. Download the pretrained model here or Baiduyun(password is e2ec).

  2. Prepared the KINS dataset according to the INSTALL.md.

  3. Test:

    We will soon release the code for e2ec with multi component detection. Currently only supported for testing e2ec performance on cityscapes dataset.

    # testing segmentation accuracy on Cityscapes with coco evaluator
    python test.py cityscapesCoco --checkpoint /path/to/model_cityscapes.pth
    # with cityscapes official evaluator
    python test.py cityscapes --checkpoint /path/to/model_cityscapes.pth
    # testing the detection accuracy
    python test.py cityscapesCoco \
    --checkpoint /path/to/model_cityscapes.pth --eval bbox
    # testing the speed
    python test.py cityscapesCoco \
    --checkpoint /path/to/model_cityscapes.pth --type speed
    # testing the contours of specified stage(init/coarse/final/final-dml)
    python test.py cityscapesCoco \
    --checkpoint /path/to/model_cityscapes.pth --stage coarse
    # testing on test set, run and submit the result file
    python test.py cityscapes --checkpoint /path/to/model_cityscapes.pth \
    --dataset cityscapes_test
    

Evaluate boundary AP

  1. Install the Boundary IOU API according boundary iou.

  2. Testing segmentation accuracy with coco evaluator.

  3. Using offline evaluation pipeline.

    python /path/to/boundary_iou_api/tools/coco_instance_evaluation.py \
        --gt-json-file /path/to/annotation_file.json \
        --dt-json-file data/result/result.json \
        --iou-type boundary
    

Visualization

  1. Download the pretrained model.

  2. Visualize:

    # inference and visualize the images with coco pretrained model
    python visualize.py coco /path/to/images \
    --checkpoint /path/to/model_coco.pth --with_nms True
    # you can using other pretrained model, such as cityscapes 
    python visualize.py cityscapesCoco /path/to/images \
    --checkpoint /path/to/model_cityscapes.pth
    # if you want to save the visualisation, please specify --output_dir
    python visualize.py coco /path/to/images \
    --checkpoint /path/to/model_coco.pth --with_nms True \
    --output_dir /path/to/output_dir
    # visualize the results at different stage
    python visualize.py coco /path/to/images \
    --checkpoint /path/to/model_coco.pth --with_nms True --stage coarse
    # you can reset the score threshold, default is 0.3
    python visualize.py coco /path/to/images \
    --checkpoint /path/to/model_coco.pth --with_nms True --ct_score 0.1
    # if you want to filter some of the jaggedness caused by dml 
    # please using post_process
    python visualize.py coco /path/to/images \
    --checkpoint /path/to/model_coco.pth --with_nms True \
    --with_post_process True
    

Training

We have released the code for multi GPU training with ddp.

Training with multi GPUS

CUDA_VISIBLE_DEVICES=${gpu_ids} python -m torch.distributed.launch \
--nproc_per_node ${gpu_nums} \
train_net_ddp.py \
--config_file ${dataset} \
--bs ${bs_per_gpu} \
--gpus ${gpu_nums}
# the example of training sbd dataset using 2 gpus
CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch \
--nproc_per_node 2 \
train_net_ddp.py \
--config_file sbd \
--bs 12 \
--gpus 2

Training on SBD

python train_net.py sbd --bs $batch_size
# if you do not want to use dinamic matching loss (significantly improves 
# contour detail but introduces jaggedness), please set --dml as False
python train_net.py sbd --bs $batch_size --dml False

Training on KINS

python train_net.py kitti --bs $batch_size

Training on Cityscapes

python train_net.py cityscapesCoco --bs $batch_size

Training on COCO

In fact it is possible to achieve the same accuracy without training so many epochs.

# first to train with adam
python train_net.py coco --bs $batch_size
# then finetune with sgd
python train_net.py coco_finetune --bs $batch_size \
--type finetune --checkpoint data/model/139.pth

Training on the other dataset

If the annotations is in coco style:

  1. Add dataset information to dataset/info.py.

  2. Modify the configs/coco.py, reset the train.dataset , model.heads['ct_hm'] and test.dataset. Maybe you also need to change the train.epochs, train.optimizer['milestones'] and so on.

  3. Train the network.

    python train_net.py coco --bs $batch_size
    

If the annotations is not in coco style:

  1. Prepare dataset/train/your_dataset.py and dataset/test/your_dataset.py by referring to dataset/train/base.py and dataset/test/base.py.

  2. Prepare evaluator/your_dataset/snake.py by referring to evaluator/coco/snake.py.

  3. Prepare configs/your_dataset.py and by referring to configs/base.py.

  4. Train the network.

    python train_net.py your_dataset --bs $batch_size
    

Citation

If you find this project helpful for your research, please consider citing using BibTeX below:

@inproceedings{zhang2022e2ec,
  title={E2EC: An End-to-End Contour-based Method for High-Quality High-Speed Instance Segmentation},
  author={Zhang, Tao and Wei, Shiqing and Ji, Shunping},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={4443--4452},
  year={2022}
}

Acknowledgement

Code is largely based on Deep Snake. Thanks for their wonderful works.

About

E2EC: An End-to-End Contour-based Method for High-Quality High-Speed Instance Segmentation

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published