This is a partial implementation of the paper : Mousavian, Arsalan, et al. "3D Bounding Box Estimation Using Deep Learning and Geometry." arXiv preprint arXiv:1612.00496 (2016).
The aim of this project is to predict the size of the bounding box and orientation of the object in 3D space from a single two dimensional image. The paper implements a 3D location estimation algorithm as well which we haven't yet implemented. Although, it'll be a good addition. Moreover, we consider only one bin for orientation while the paper suggests two.
- TensorFlow 1.0
- Numpy
- OpenCV 2
- Python 2.7
- tqdm
- Git clone this project : git clone https://github.com/shashwat14/Multibin.git
- Download the KITTI object detection dataset and save it within the Multibin folder or save it somewhere else and make changes as mentioned in point 3.
- Open helper.py and edit the following paths and make sure the path names are correct : path = /path/to/Multibin/ train_path = path + 'training_data/' train_images_path = train_path + 'images_1/' train_labels_path = train_path + 'labels_1/'
- Download the weights file (vgg16_weights.npz) and place in Multibin directory. https://www.cs.toronto.edu/~frossard/vgg16/vgg16_weights.npz
- python main.py
- python train.py
- python test.py