SLICE - Semi-supervised Learning approach for Image segmentation of Crop field boundariEs

AIM: Produce accurate field boundary segmentation maps in regions with little to no ground truth labels.

METHOD: Semi-supervised Instance segmentation

TRACKER: Notion link

---

Usage

• Perform a Recursive git clone

  git clone --recursive [email protected]:kerner-lab/slice.git 
  

• Change a code in Transformers library under models folder

  • Path slice/models/transformers/src/transformers/models/sam/image_processing_sam.py

  • Modification

  Line 938 :  (Error)
  crop_boxes = crop_boxes.astype(np.float32)
  
  should work with
  crop_boxes = np.array(crop_boxes, dtype=np.float32)
  

  • Perform installation for Transformers library

  pip install -e .
  

---

Guidelines

• Following semantic versioning with Major.Minor.Patch
• Every small changes will come under patch version. (bugfix, bug)
• Every feature introduction will change the minor version.
• Every interface logic change will change the major version.
• Example of the versioning
  • 2.0.005
    • Major version - 2
    • Minor Version - 0
    • Patch Version - 005
  • Always create a new branch for a feature introduction and raise a pull request accordingly.
  • Releases will be tagged either Pre Release or Latest Release from the main branch only.
  • Git commits will have the changelogs description as the commit message. e.g. 2.0.005 (feature+update)
  • Definition of keywords
    • feature - Any introduction of a new feature
    • bug - With the commited change the system is in a buggy state.
    • bugfix - for any patches applied over a bug.
    • update - general updates like updating readme. (this won't increment any version numbers)
    • experimental - This stays out of the main branch unless the experiment is solidified to create a feature out of it.

---

Changelogs

0.0.1 (feature + bugfix)

• Added Segment Anything Project as a submodule.
• Fixed and issue with the setup_sol.sh file.

0.0.2 (feature)

• Added code to extract tf record.
• Added code to generate pseudo masks for the image dataset using SAM.

0.0.3 (bug + bugfix)

• The conversion code for tf record is not working properly. -> BUG
• Added script for downloading the model checkpoints.
• Made path changes in generate_pseudo_masks.sh file.

0.0.4 (feature + bugfix)

• Conversion for TFrecord
• Extract images script for inference
• Output visualization script .ipynb

0.0.5 (feature)

• Working with the combine patches for display. The patches(masks) are generated as separate files (instances)

0.0.6 (feature)

• Dataset creating script added to data.ipynb

0.0.7 (feature + bugfix + update)

• The Dataset loader is now available.
• A Visualization notebook is available. See infer_and_eval.ipynb

0.1.0 (feature)

• Added Transformers Library from HuggingFace
• Recreated Inference Pipeline for better understanding and decoupling it from the HuggingFace internal Pipeline.
• WIP for training script.