In this work,we are aiming at predicting instance-wise depth via self-supervised learning mechanism. The introduction of instance segmentation network is essential to dealing with dynamic objects, in turn a high precise depth map is also helpful for object detection/instance segmentation tasks.
generally speaking, any instance seg networks could be inserted into our system, however, here is an open source toolbox including a series of instance segmentation algorithms. Up to now, These algorithms have achieved better performance than most SOTA methods on COCO dataset.
AdelaiDet consists of the following algorithms:
Initialy, I was planning to develop my own system based on the mentioned open-source framework. However, releasing their implementations may need a longer time than we expect. Therefore, I resort to CenterMask, which is also a one-stage instance sgementation network based on FCOS. The original implementation is based on maskrcnn-benchmark, but I reformulate their code, add some new features, and will be opened at here.
It should be mentioned that AdelaiDet is extended from Detectron2, so please install Detectron2 following the official guide: INSTALL.md. Then build AdelaiDet with:
git clone https://github.com/aim-uofa/adet.git
cd adet
python setup.py build develop
- Pick a model and its config file, for example,
fcos_R_50_1x.yaml. - Download the model
wget https://cloudstor.aarnet.edu.au/plus/s/glqFc13cCoEyHYy/download -O fcos_R_50_1x.pth - Run the demo with
python demo/demo.py \
--config-file configs/FCOS-Detection/R_50_1x.yaml \
--input input1.jpg input2.jpg \
--opts MODEL.WEIGHTS fcos_R_50_1x.pth
To train a model with "train_net.py", first setup the corresponding datasets following datasets/README.md, then run:
python tools/train_net.py \
--config-file configs/FCOS-Detection/R_50_1x.yaml \
--num-gpus 8 \
OUTPUT_DIR training_dir/fcos_R_50_1x
The configs are made for 8-GPU training. To train on another number of GPUs, change the num-gpus.