Image Segmentation Instrumentation

This repository contains instrumentation to utilize a Docker image that contains a popular segmentation model detectron2.

We utilize a forked version of detectron2.

Our instrumentation does the following:

• clone and set up the base repository.
• build image, cache pre-trained models, and produce predictions on this input image.

We depend on docker version >= 19.03.

Execution

We use a Makefile for instrumentation of a popular open source model detectron2.

Obtain detectron:

make clone

This downloads detectron.

make make_dir

This sets up the expected directory structure.

We use a sample image to demonstrate the feature. First obtain the image using:

wget http://images.cocodataset.org/val2017/000000439715.jpg -O input.jpg

Now, we set up the input image for segmentation.

make input_image img=input.jpg

Now we run segmentation:

make run img=input.jpg

Which will store the visualization at outputs/input.jpg and the torch tensors at outputs/input.jpgpredictions.

You may of course parse the tensors directly for use. But you may also utilize the interactive setting to inspect models. The interactive setting drops into a Docker container that contains the detectron repository. From there, models can be executed following the instructions at detectrons2.

The model's prediction object contains the key instances which contains a list of Instances objects:

Instances(
	num_instances=1,
	image_height=480,
	image_width=640,
	fields=[
		pred_boxes: Boxes(...),                      # box coordinates
		scores: tensor(...),                         #
		pred_classes: tensor([17], device='cuda:0'), # prediction
		pred_masks: tensor(),                        # prediction masks
	)

The sample output visualized:

Utilize the interactive setting to inspect models. The interactive setting drops into a docker container that contains the detectron repo. From there, models can be executed following the instructions at detectrons2.

make interactive

Inside the image, detectron2 is cloned, and available for execution. We can utilize it for inspecting the dumped predictions.