Awesome

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

• The pretrained EfficientNet weights on imagenet are downloaded from Callidior/keras-applications/releases
• The pretrained EfficientDet weights on coco are converted from the official release google/automl.

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.