MobileDet

MobileNet and YOLOv2 based real-time Person and Vehicle detector. The final verion of model is called: YOLOv2-MobileNet_Shallow_3Scales This model can run up to 10fps on NVIDIA TX2 platform

Dependencies

Keras
Tensorflow
scikit-learn
Numpy
h5py  
Pillow  
Python 3
opencv-python

UAV123 dataset

Test with UAV123 video sequences

• 0. Download the UAV123 dataset zip file from link. It is about 4.4G. Unzip it into /data/UAV123/ folder. Download the pre-processed UAV123 dataset from link and placed it under '~/data' folder.
• 1. Download weights from this google drive link to UAV123 weights and place the unzipped folder to 'MobileDet/weights_uav123'
• 2. Modify the cfg.py as below to configure the model as YOLOv2-MobileNet_Shallow_3Scales

FEATURE_EXTRACTOR = 'mobilenet'
SHALLOW_DETECTOR = True
USE_X0_FEATURE = True

• 3. Run below sample script to use the model to detect objects from UAV123 images

python test_yolo.py  -m weights_uav123/mobilenet_s3_best.TrueTrue.h5  -t ~/data/UAV123/UAV123_10fps/data_seq/UAV123_10fps/bike3 -o  ~/Videos/bike3 -iou 0.6 -s 0.6

• 4. Run below sample script to use calculate the recall and precision of Test group UAV123 dataset

python recall_precision.py  -m weights_uav123/mobilenet_s3_best.TrueTrue.h5 -d ~/data/uav123.hdf5 -a model_data/uav123_anchors.txt -iou 0.6 -s 0.6

Retrain with other dataset

Pre-process dataset

• 1. To retrain with other dataset, it is required to write a parse script to convert the dataset into HDF5 format. This project has python script for dataset: UAV123, VOC and Okutama Action datasets
• 2. After creating the HDF5 dataset, it is required to use the 'anchor_boxes.py' to run k-means method to generate a few prior anchor boxes for this dataset
• 3. After above two steps, use 'retrain_yolo.py' to retrain the new dataset

YAD2K

This project is based on: https://github.com/allanzelener/YAD2K