Keras Implementation of Single Shot MultiBox Detector

A pure Tensorflow+Keras Implementation of SSD (Single Shot MultiBox Detector) using different backbones of Efficientnet on the PASCAL_VOC dataset.

Dependencies

1. Python 3.6+
2. Tensorflow 2.2.0 (For GPU, use Tensorflow-GPU 2.2.0)
3. Tensorflow_Datasets 3.0.0+
4. Efficientnet

To install these dependencies for CPU, run

pip install -r requirements.txt

(For GPU support, change line tensorflow==2.2.0 to tensorflow-gpu==2.2.0 in requirements.txt)

Test on your own images

Add your images to the inputs/ folder and then run

python predict.py

A pretrained model with EfficientNetB0 backbone will load and run on all images in the inputs/ folder. The results can be found in the outputs/ folder.

NOTE:

To train or evaluate the model would require downloading the PASCAL VOC dataset and converting it into tfrecords format by tensorflow_datasets module. This is done automatically in train.py or eval.py but requires a considerable amount of time on the first run.

Evaluate Model

In eval.py, change the checkpoint_filepath variable to your trained model weights and make sure the base model i.e. MODEL_NAME is set according to your architecture ('B0', 'B1'...etc.). Then run

python eval.py

PASCAL_VOC evaluation will be performed on the VOC2007 test dataset.

Train Model

To train from scratch, run

python train.py

If you want to continue training from a checkpoint set checkpoint_filepath and set base_lr accordingly.

Pretrained Models

Original Paper

Backbone	mAP	Model Size
VGG16	77.2	~200 MB

In this Repo

Backbone	mAP	Model Size	Download weights
EfficientNet-B0	74.3	24MB	model
EfficientNet-B3	76.3	49MB	model
EfficientNet-B5	77.5	117MB	model

EfficientNetB0 SSD (300 x 300)

****************************************************************************************************
Average Precisions
{'aeroplane': 0.7800157231291908,
 'bicycle': 0.830080097782427,
 'bird': 0.6972303731738765,
 'boat': 0.6874695492244346,
 'bottle': 0.44450301285192956,
 'bus': 0.7784750741179701,
 'car': 0.8118978656479396,
 'cat': 0.8710704906835629,
 'chair': 0.6031652990795755,
 'cow': 0.7280159632333296,
 'diningtable': 0.7498881454649887,
 'dog': 0.8379743383251255,
 'horse': 0.8337120771431471,
 'motorbike': 0.8194014393196231,
 'person': 0.774534918174515,
 'pottedplant': 0.5322147125048592,
 'sheep': 0.7044521803545073,
 'sofa': 0.7880689464995533,
 'train': 0.8751272977754305,
 'tvmonitor': 0.7190302090784884}
****************************************************************************************************
Mean Average Precision: 0.7433163856782238

EfficientNetB3 SSD (300 x 300)

****************************************************************************************************
Average Precisions
{'aeroplane': 0.8138464069658522,
 'bicycle': 0.8495817279095099,
 'bird': 0.7411210589421086,
 'boat': 0.6953653156671167,
 'bottle': 0.44501229699670486,
 'bus': 0.8169842240921157,
 'car': 0.8348353061776034,
 'cat': 0.867478526227363,
 'chair': 0.6359432618499943,
 'cow': 0.7347537326752558,
 'diningtable': 0.7556188461308526,
 'dog': 0.8453690345281261,
 'horse': 0.8431191113749101,
 'motorbike': 0.8414147988740155,
 'person': 0.8000506418143006,
 'pottedplant': 0.5718438091150522,
 'sheep': 0.7248609390832621,
 'sofa': 0.8101149566442967,
 'train': 0.8836348165299509,
 'tvmonitor': 0.7443892209541666}
****************************************************************************************************
Mean Average Precision: 0.762766901627628

EfficientNetB5 SSD (300 x 300)

****************************************************************************************************
Average Precisions
{'aeroplane': 0.8402582577317493,
 'bicycle': 0.8611873325373057,
 'bird': 0.7579335634733774,
 'boat': 0.7394833127167081,
 'bottle': 0.47933192495395077,
 'bus': 0.8181464002356922,
 'car': 0.8443131628054835,
 'cat': 0.8799997608338624,
 'chair': 0.6257301589722992,
 'cow': 0.7870176718398668,
 'diningtable': 0.7662971953267264,
 'dog': 0.8378183820936017,
 'horse': 0.8493214449256444,
 'motorbike': 0.8534871863365513,
 'person': 0.8214934259604633,
 'pottedplant': 0.5706558290952545,
 'sheep': 0.7471660716329726,
 'sofa': 0.7908692690722835,
 'train': 0.8834500074371311,
 'tvmonitor': 0.7397631809290843}
****************************************************************************************************
Mean Average Precision: 0.7746861769455005

Citation

If you find this repo useful in your work and build something interesting, do cite to let me know about it!

@software{swapnil_gupta_2021_4775343,
  author       = {Swapnil Gupta},
  title        = {SSD implementation in Keras with EfficientNet},
  month        = may,
  year         = 2021,
  publisher    = {Zenodo},
  version      = {1.1},
  doi          = {10.5281/zenodo.4775343},
  url          = {https://doi.org/10.5281/zenodo.4775343}
}

References

1. https://github.com/qfgaohao/pytorch-ssd
2. https://github.com/lufficc/SSD
3. https://github.com/mvoelk/ssd_detectors