EfficientDet: Scalable and Efficient Object Detection

PyTorch Implementation of the state-of-the-art object detection architecture EfficientDet

https://arxiv.org/abs/1911.09070

Motivation

As of the time I started working on this project, there was no PyTorch implementation on GitHub that would match the original paper in the number of the model's parameters. All of the existed repositories altered a lot from the recently published TensorFlow implementation by Brain Team (e.g. changing strides in the backbone, missing batch normalization layers, no 'same' padding strategy in pooling layers, differing training hyper-parameters, not using Exponential Moving Average Decay, and others). Here is my attempt to reproduce EfficientDet in PyTorch. My end goal is to reproduce training cycle from the original paper and achieve nearly same results.

Notes on Implementation

Alternatively to the TensorFlow implementation, I got rid of the useless biases in convolutional layers followed by batch normalization, which resulted in parameters reduction.

Model Zoo

Usage

Train from scratch

Download COCO2017 Train & Val Sets

wget http://images.cocodataset.org/zips/train2017.zip
unzip train2017.zip && mv train2017 data/coco/train2017 && rm train2017.zip

wget http://images.cocodataset.org/zips/val2017.zip
unzip val2017.zip && mv val2017 data/coco/val2017 && rm val2017.zip

wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
unzip annotations_trainval2017.zip && mv annotations data/coco && rm annotations_trainval2017.zip

Run Script

python main.py -mode 'trainval' -model 'efficientdet-d{}'

COCO Evaluation

Download COCO2017 Val Set

wget http://images.cocodataset.org/zips/val2017.zip
unzip val2017.zip && mv val2017 data/coco/val2017 && rm val2017.zip

wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
unzip annotations_trainval2017.zip && mv annotations data/coco && rm annotations_trainval2017.zip

Run Script

python main.py -mode eval -model efficientdet-d{} --pretrained

RoadMap

• Model Architecture that would match the original paper
• COCO val script
• port weights from TensorFlow
• COCO train script
• Reproduce results from the paper
• Pre-trained weights release

References

• [1] Mingxing Tan, Ruoming Pang, Quoc V. Le. EfficientDet: Scalable and Efficient Object Detection
• [2] EfficientDet implementation in TensorFlow by Google AutoML
• [3] PyTorch EfficientNet implementation by @lukemelas