Awesome

Old Motion R-CNN

First attempt - switched to tensorflow object_detection. See the motion-rcnn repo.

Requirements

• tensorflow (>= 1.2.0) with GPU support. For best performance, i highly recommend building from source.
• pip install pillow matplotlib opencv-python easydict cython tqdm

Setup

• create ./out directory
• copy env_template/env.yml to out/env.yml and adapt for your machine setup
• download http://download.tensorflow.org/models/resnet_v1_50_2016_08_28.tar.gz and unzip to ./data/models/
• go to ./lib and run make
• run tools/create_tfrecords.py with each --dataset/--split combination you need

Usage

• run python tools/trainval.py for training
• run python tools/test.py for testing

Unit Tests

• run python test/cityscapes.py to visualize the cityscapes dataset
• run python test/anchors.py to visualize anchors for different levels

Visualizations are written to out/tests.

Acknowledgments

• The code in lib/nms and lib/boxes is taken without changes from py-faster-rcnn.
• The tensorflow code in lib/nets/resnet_v1.py and lib/nets/network.py is based on tf-faster-rcnn.
• The files implementing common Faster R-CNN layers in lib/layers are based on py-faster-rcnn and include small modifications from tf-faster-rcnn.
• The code in lib/datasets/cityscapes/cityscapesscripts is adapted from cityscapesScripts.

License

See LICENSE for details.