This is a Work In Progress Tensorflow implementation of "End-to-End Learning of Geometry and Context for Deep Stereo Regression"
Below picture shows the training samples after 2 epochs in Tensorboard over Sceneflow dataset. It's huge dataset and my potato laptop can't handle more epochs.
Normally here I'd give instruction to install but Tensorflow installation is complicated. So I can only tell you that this work is done in tensorflow 2.2.2, Ubuntu 20.04, CUDA 10.1, CUDNN 7.6.5
See https://www.tensorflow.org/install/source#gpu for more info about Tensorflow version related to CUDA+CUDNN
python train.py --cfg config/potato_laptop.yaml
TODO:
What's working?
- Full GC Net network
- Cost Volume -> Provide better information of disparity
- Soft Argmin / Soft Argmax -> Provide subpixel disparity
- Customizable Config through YAML
- Preprocessing
- Random Crop based on the network training height and width -> Can handle more epochs without overfitting (but my potato laptop can't handle it)
- Tensorboard images and loss logging
- Checkpoint save file