This is a Mask R-CNN implementation with MobileNet V1/V2 as Backbone architecture to be finally able to deploy it on mobile devices such as the Nvidia Jetson TX2.
The major changes to the original matterport project are:
- Add Mobilenet V1 and V2 as backbone options (besides ResNet 50 and 101) + dependencies in the model
- Make the whole project py2 / py3 compatible (original only works on py3)
- Investigate Training Setup for Mobilenet V1 and implement it in
coco_train.py
- Add a Speedhack to mold /unmold image functions
- Make the project lean and focused on COCO + direct training on passed class names (IDs before)
- Inclue more speed up options to the Model (Light-Head RCNN)
- Release a trained Mobile_Mask_RCNN Model
- install required packages (mostly over pip)
- clone this repository
- download and setup the COCO Dataset:
setup_coco.py
- inside
coco.py
subclassConfig
(defined inconfig.py
) and change model params to your needs - train
mobile mask r-cnn
on COCO with:train_coco.py
- evaluate your trained model with:
eval_coco.py
- do both interactively with the notebook
train_coco.ipynb
- if you face killed kernels due to memory errors, use
bash train.sh
for infinite training - visualize / control training with tensorboard:
cd
into your current log dir and run:
tensorboard --logdir="$(pwd)"
- inspect your model with
notebooks/
:
inspect_data.ipynb
,inspect_model.ipynb
,inspect_weights.ipynb
,detection_demo.ipynb
- convert keras h5 to tensorflow .pb model file, in
notebooks/
run:
export_model.ipynb
Mobile Mask R-CNN trained on 512x512 input size
- 100 Proposals: 0.22 mAP (VOC) @ 250ms
- 1000 Proposals: 0.25 mAP (VOC) @ 330ms
- numpy
- scipy
- Pillow
- cython
- matplotlib
- scikit-image
- tensorflow>=1.3.0
- keras>=2.1.5
- opencv-python
- h5py
- imgaug
- IPython[all]
- pycocotools