ajaichemmanam / CenterSeg

This project uses Centernet and Conditional Convolutions for Instance Segmentation

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

CenterSeg

This repo uses Centernet and Conditional Convolutions for Instance Segmentation

Objects as Points,
CondInst: Conditional Convolutions for Instance Segmentation

Result

These results are taken for CenterSeg model trained for 101 epochs

type AP AP50 AP75 APs APm APl
box 0.278 0.430 0.297 0.129 0.305 0.382
mask 0.226 0.387 0.227 0.078 0.253 0.340
type AR AR50 AR75 ARs ARm ARl
box 0.275 0.455 0.480 0.265 0.510 0.674
mask 0.235 0.369 0.385 0.170 0.418 0.585

CenterPoseSeg model not trained yet

Installation

This repo supports both CPU and GPU Training and Inference.

git clone --recurse-submodules https://github.com/ajaichemmanam/CenterSeg.git

pip3 install -r requirements.txt

Compile DCN

cd src/lib/models/networks/DCNv2/

python3 setup.py build develop

Compile NMS

cd src/lib/external

python3 setup.py build_ext --inplace

Pre-Trained Models

Pre-Release : Google Drive

Download the most recent model (model_last_e101.pth), copy to exp/ctseg/coco_dla_1x/

Rename as model_last.pth

python3 demo.py ctseg --exp_id coco_dla_1x --keep_res --resume --demo ../data/coco/val2017

Note: Model is not completely trained (101 Epochs only). Will update later.

Training

For GPU
python3 main.py ctseg --exp_id coco_dla_1x --batch_size 10 --master_batch 5 --lr 1.25e-4 --gpus 0 --num_workers 4
FOR CPU
python3 main.py ctseg --exp_id coco_dla_1x --batch_size 2 --master_batch -1 --lr 1.25e-4 --gpus -1 --num_workers 4

Testing

python3 test.py ctseg --exp_id coco_dla_1x --keep_res --resume

License

CenterSeg is released under the MIT License (refer to the LICENSE file for details). This repo contains code borrowed from multiple sources. Please see their respective licenses.

Credits

https://github.com/xingyizhou

https://github.com/Epiphqny

https://github.com/CaoWGG

About

This project uses Centernet and Conditional Convolutions for Instance Segmentation

License:MIT License


Languages

Language:Python 83.1%Language:C++ 13.8%Language:Shell 2.1%Language:Cuda 1.0%Language:Makefile 0.0%