Skip to content
/ cseg Public

Context-aware open-vocabulary semantic segmentation (adapted from ov-seg)

Notifications You must be signed in to change notification settings

noellelaw/cseg

Repository files navigation

cseg

Context-aware open-vocabulary semantic segmentation (adapted from ov-seg)

Data Preparation

Please see datasets preparation.

Getting Started

Evaluation on the KITTI Dataset with OVSeg Pretrained Weights: (Swin-Base + CLIP-ViT-L/14)

  • Download cseg_verification.ipynb, or view colab file directly
  • Upload ovseg_swinbase_vitL14_ft_mpt.pth to accessable folder, preferably a google drive folder so that you don't have to re-upload every time. Edit the model_weights field under Build Model in the ipynb file to the path where ovseg_swinbase_vitL14_ft_mpt.pth is stored.
  • Place the KITTI Test images to an easily accessible place in drive. Edit the data_fldr field under Read in KITTI test data from drive in the ipynb file to this path.

Training on COCO or ADE20k dataset:

  • The cseg_training.ipynb has the basic training commands and setup, it assumes you have gone through data preparation required for the detectron2 datasets.
  • Due to the size of the datasets, the connection between google colab and google drive can time out.
  • The /sbatch/ folder features the files required to run this in a slurm environment. It assumes that the virtual environment has all required packages installed.
  • The code for full training is in open_vocab_seg subfolder.
  • More information on training can be found here.

Fine-tuning the classification stage (CLIP) is adapted from OVSeg:

Required weights

About

Context-aware open-vocabulary semantic segmentation (adapted from ov-seg)

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published