Instance Segmentation Image Viewer

A Tkinter desktop app that runs a pretrained Mask R-CNN on an image you pick and shows either its segmentation masks or its labelled bounding boxes next to the original. It is built on a small Python package, newmypackage, with a dataset loader, image transforms, a model wrapper and plotting. This was a Software Engineering Lab assignment at IIT Kharagpur.

Bounding boxes	Segmentation

Features

• Open an image with the . . . . . button. The model runs straight away, and the selected path appears in the text box.
• Choose a view from the drop-down, Segmentation or Bounding-box, then click Go to switch views without re-running the model.
• Shows the three most confident detections. Boxes are labelled with their COCO class names, and the masks of the top three detections are combined.
• Shows a message if Go is clicked before an image is chosen, or if the chosen image is not in the annotation file.

Getting started

Requirements: Python 3 with Tkinter, plus the packages in requirements.txt: PyTorch, torchvision, NumPy, Pillow and Matplotlib.

pip install -r requirements.txt
python ImageViewerGUI.py        # run from the repository root

The first run downloads the pretrained Mask R-CNN weights (about 170 MB). If Python reports No module named '_tkinter', install Tk support for your Python, for example brew install python-tk on macOS or sudo apt install python3-tk on Debian/Ubuntu.

The app reads data/annotations.jsonl and the images in data/imgs/, so choose images from that folder. It writes its rendered views to output/.

The newmypackage package

Module	Contents
model.py	InstanceSegmentationModel, a wrapper around torchvision's maskrcnn_resnet50_fpn (COCO-pretrained). Takes a (3, H, W) image in [0, 1] and returns boxes, masks, class names and scores.
data/dataset.py	Dataset, which reads a JSON Lines annotation file and returns the image, the segmentation PNG and the ground-truth boxes for an index, applying any transforms in order.
data/transforms/	BlurImage (Gaussian), CropImage (centre or random), FlipImage (horizontal or vertical), RescaleImage (to a fixed size or by the shorter side), RotateImage
analysis/visualize.py	plot_visualization, which draws the original image next to the masks or the labelled boxes and saves the figure

The package can be installed on its own:

pip install .

from newmypackage.data.dataset import Dataset
from newmypackage.data.transforms.flip import FlipImage
from newmypackage.data.transforms.blur import BlurImage

ds = Dataset('data/annotations.jsonl', transforms=[FlipImage('horizontal'), BlurImage(2)])
item = ds[0]          # {'image': (H, W, 3) array, 'gt_png_ann': ..., 'gt_bboxes': [[class, x, y, w, h], ...]}

Data

data/ holds 10 sample COCO images with their segmentation PNGs. data/README.md describes the annotation format.

Author

Deepiha S (20CS30015), Software Engineering Laboratory, IIT Kharagpur.