Skip to content

EfficientDet (Scalable and Efficient Object Detection) implementation in Keras and Tensorflow

License

Notifications You must be signed in to change notification settings

kirumang/EfficientDet

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

57 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

EfficientDet

This is an implementation of EfficientDet for object detection on Keras and Tensorflow. The project is based on the official implementation google/automl, fizyr/keras-retinanet and the qubvel/efficientnet.

About pretrained weights

Thanks for their hard work. This project is released under the Apache License. Please take their licenses into consideration too when use this project.

Updates

  • [03/21/2020] Synchronize with the official implementation. google/automl
  • [03/05/2020] Anchor free version. The accuracy is a little lower, but it's faster and smaller.For details, please refer to xuannianz/SAPD
  • [02/20/2020] Support quadrangle detection. For details, please refer to README_quad

Train

build dataset

  1. Pascal VOC
    • Download VOC2007 and VOC2012, copy all image files from VOC2007 to VOC2012.
    • Append VOC2007 train.txt to VOC2012 trainval.txt.
    • Overwrite VOC2012 val.txt by VOC2007 val.txt.
  2. MSCOCO 2017
    • Download images and annotations of coco 2017
    • Copy all images into datasets/coco/images, all annotations into datasets/coco/annotations
  3. Other types please refer to fizyr/keras-retinanet)

train

  • STEP1: python3 train.py --snapshot imagenet --phi {0, 1, 2, 3, 4, 5, 6} --gpu 0 --random-transform --compute-val-loss --freeze-backbone --batch-size 32 --steps 1000 pascal|coco datasets/VOC2012|datasets/coco to start training. The init lr is 1e-3.
  • STEP2: python3 train.py --snapshot xxx.h5 --phi {0, 1, 2, 3, 4, 5, 6} --gpu 0 --random-transform --compute-val-loss --freeze-bn --batch-size 4 --steps 10000 pascal|coco datasets/VOC2012|datasets/coco to start training when val mAP can not increase during STEP1. The init lr is 1e-4 and decays to 1e-5 when val mAP keeps dropping down.

Evaluate

  1. PASCAL VOC

    • python3 eval/common.py to evaluate pascal model by specifying model path there.
    • The best evaluation results (score_threshold=0.01, mAP50) on VOC2007 test are:
    phi 0 1
    w/o weighted 0.8029
    w/ weighted 0.7892
  2. MSCOCO

    • python3 eval/coco.py to evaluate coco model by specifying model path there.
    phi mAP
    0 0.334 weights, results
    1 0.393 weights, results
    2 0.424 weights, results
    3 0.454 weights, results
    4 0.483 weights, results

Test

python3 inference.py to test your image by specifying image path and model path there.

image1

About

EfficientDet (Scalable and Efficient Object Detection) implementation in Keras and Tensorflow

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%