imbibekk / ssd_tf2

tf2x implementation of SSD

SSD-tf2.0

tf2.0 Implementation of SSD

This repo is inspired by SSD Pytorch and can be seen as its porting in tf2.0

Installation

All requirements should be detailed in requirements.txt. Using Anaconda is strongly recommended.

conda create -n ssd_tf2 python=3.6
source activate ssd_tf2
pip install -r requirements.txt

Prepare dataset

You need VOC 2007 and VOC 2012 data. If you don't have already, you can download it by

$ sh download_voc.sh

After downloading and unzipping, the data directory should look like this:

data
  +- pascal_voc
    +- VOCdevkit
      +- VOC2007
      +- VOC2012

Training

Training can be done by using the config file in configs folder.

python main.py --config configs/vgg_ssd300_voc0712.yaml

log.txt file is attached for your reference

Evaluate

For evaluating the trained model. Model weights can be downloaded via this link

python main.py --config configs/vgg_ssd300_voc0712.yaml --test True CKPT model_weights.h5

AP_aeroplane : 0.8374984881786407
AP_bicycle : 0.8464588448722019
AP_bird : 0.7590097561192465
AP_boat : 0.7142750067666049
AP_bottle : 0.5092045441421713
AP_bus : 0.8577446126445106
AP_car : 0.8591203417934831
AP_cat : 0.885377562894929
AP_chair : 0.6205903576229672
AP_cow : 0.8181410670105621
AP_diningtable : 0.7646054306807052
AP_dog : 0.8467388955700131
AP_horse : 0.8672697782501421
AP_motorbike : 0.8306251242356146
AP_person : 0.793392521910993
AP_pottedplant : 0.5204622608984472
AP_sheep : 0.7652274001107799
AP_sofa : 0.8065167013126614
AP_train : 0.8623668124696708
AP_tvmonitor : 0.7666016840177408
mAP : 0.7765613595751042