Conversion made by Rafael Berral-Soler, based on the original work by Pablo Medina-Suarez and Manuel J. Marin-Jimenez.
This model relies on the Keras implementation of Single-Shot Multibox Detector by Pierluigi Ferrari found here.
-
In order to use this model, you MUST import models/keras_ssd512.py, bounding_box_utils and keras_layers from its repository.
-
Because of the size of the converted model, to clone the repository git-lfs is needed; the model can also be converted from the original MatConvNet model using convert_ssd_512.py.
Instructions for setting up git-lfs alongside git (from ssd_people repository):
Install git: sudo apt-get install git Install git-lfs: sudo apt-get install git-lfs Set up git-lfs: git lfs install Clone ssd_people_keras from GitHub using the method of your choice: git clone https://github.com/AVAuco/ssd_people_keras.git (HTTPS) git clone git@github.com:AVAuco/ssd_people_keras.git (SSH)
-
Script convert_ssd_512.py may be able to convert other SSD-512 models, provided they use the same MatConvNet toolbox used in the original work. Slight modifications to the script and the layer mapping (layers.csv) could make possible to convert also SSD-256 models.
-
Detections using the converted model could not match detections obtained with the original MatConvNet implementation. Adjusting confidence threshold should improve performance.
-
In order to use the model and code included in this repository, it may be useful to update your PYTHONPATH. Provided you cloned this repository at ~/libs/ssd_keras:
export PYTHONPATH=$PYTHONPATH:~/libs/ssd_keras/:~/libs/ssd_keras/models/
In order to run demo.py:
- Python packages: numpy, imageio, matplotlib, keras.
- Keras SSD implementation:
- models/ssd512.py
- bounding_box_utils
- keras_layers
In order to use convert_ssd_512.py:
- Python packages: numpy, pandas, keras, scipy.
- loadmat_stackoverflow.py, code obtained from here (accessed on April 11, 2019).
- Keras SSD implementation:
- models/ssd512.py
- bounding_box_utils
- keras_layers
Picture used in demo by Ross Broadstock. Licensed under CC BY 2.0 license.