Skip to content

Panoptic Segmentation - Internship @ AI without Borders 2019

Notifications You must be signed in to change notification settings

bsridatta/Panoptic-Segmentaion-AIWB

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

28 Commits
 
 

Repository files navigation

Panoptic Segmentation

Internship @ AI Without Borders

Week - 1

Literature Study - Panoptic Segmentation

Paper bottom line Constraints Code Dataset used
Panoptic Segmentation by FAIR PQ - Considers segmentation quality and a recognition quality. PQ is insensitive to class imbalance by calculating for each class independently and average over classes. Just an introduction to the task of panoptic segmentation No code Vistas, CityScapes, ADE20k
Panoptic Feature Pyramid Networks Detailed insights into the single-network architecture Requires generating a coherent scene segmentation that is rich and complete. No Code
UPSNet: A Unified Panoptic Segmentation Network 2 head network: Semantic Segmentation using Deformable Convolutions & Instance Segmentation using Mask R-CNN Complex and mutli-headed architecture, could be tricky to get it to work Uber Research COCO, CityScapes
Panoptic Segmentation with an End-to-End Cell R-CNN for Pathology Image Analysis Panoptic segmentation of various cancer cells Access to dataset. The paper does not mention details of the dataset nor if it was further modified to all pixel segmentation No code MICCAI 2017 digital pathology challenge dataset

Week - 2&3

Datasets and corresponding literature

Specifically for Panoptic Segementation

Datasets Description
COCO 2018 Things, regions (Road, grass, water)
ADE20k Things, regions and parts of things
Mapillary Vistas Street Scenes
CityScapes City/Street Scenes

Medical Image Datasets

Datasets Description
2017 MICCAI Digital Pathology Challenge dataset Segmentation of Nuclei in Images. Dataset used in a paper for panoptic segmentation.
Could not gain access to the dataset
Leukemia ALL-IDB (White) Blood cell classification. Looks suitable but have to download to know how the data is.
Multi-class artefact detection in video endoscopy Perfect dataset! But annotations are bounding box and not segementation. Have time to annotate?
Blood Cells Detection Again perfect dataset! But annotations are bounding box and not segementation. Have time to annotate?
BACH Breast Cancer Histology Images. 1 huge whole slides with multiple classes and instances but cropped into 400 which mostly 1 instance and 1 class
Melanoma, Skin Cancer (Object segmentation) Multi class segmentaion with 1 instance.
Segmentation of neuronal structures in EM stacks Home Cell boundaries hence good for normal segmentation only
Nuclear Segmentation Good segmentation dataset but only nuclei, boundaries and background. Bascially 2 classes
Medical Segmentation Decathlon Huge number of datasets with no proper description
Broad Bioimage Benchmark Collection Database - with excellent description
The cell Image Library Database
ImageJ Database - Small
List of Cancer Cell Datasets for DL Good datasets with multiclass classification but mostly for object detection (1 instance)

Week - 4

Annotation tools and label format conversions

Coco Annotator
Quick annotaion of objects using 'Magic wand' to annotate disconnected objects or with the help of an API that fetches annotations from a semi-trained network. But may be too challenging not to miss any pixels while using magic wand or brush tool.

PixelAnnotationTool
Pesudo Semi automated annotation tool which uses opencv watershed segmentation for annotaing all pixels. May be ideal for our needs

voc2coco
Convert annotation format from voc to coco. Also contains direct instructions for the BCCD dataset Blood Cells Detection mentioned above.

Downlaoding and reading annotation and trying the tools to make custom annotations

Setting up coco annotator docker. Downlaoding and coverting BCCD from VOC to coco. Takes quite some time to fully annotate each image. Very imbalanced instances. Reading artifact detection EAD dataset. Too many instances of blur and contrast which is immpossible to annotate for panoptic.

Scraping the web one last time for a good dataset

no luck

Week - 5&6

UPSNet

Going through the code of UPSNet and trying to get some predictions

Releases

No releases published

Packages

No packages published