Single image inference, Docker support, API support

[scottmartinnet]

I have completed a series of new features based on this code:

• Single image inference: split an input image into tiles, run inference on the tiles, merge and translate the resulting polygons into a single result set with pixel coordinates.
• Automatic pretrained model downloads
• Command line support
• API support
• Docker support

I am currently successfully running this project in AWS SageMaker as an inference endpoint.

[n0man-ch0wdhury]

Hey, can you help me running infer_single_image.py? I resized the image into 224 x 224, and ran the command python infer_single_image.py -e folder but no response. I have rtx 3050 4GB and i5 11th gen.

[scottmartinnet]

Make sure to provide the image as stdin, or the command will just wait for input.

Example:
cat input.png | python infer_single_image.py -e runs_share/Pix2Poly_inria_coco_224

[scottmartinnet]

Also, you don't have to resize, the code will automatically tile a large input image and assemble polygons from multiple runs.

[n0man-ch0wdhury]

@scottmartinnet Thank you for your response. It's working. I've another question. If it handles large input image, how large the input can be depending on the hardware?

[scottmartinnet]

Awesome!

The image can be arbitrarily large as long as it fits in memory and you have enough processing time to process that many tiles.

[n0man-ch0wdhury]

Okay! Is it normal to take ~10s for per 224x224 tile?

[scottmartinnet]

It all depends on your hardware.

[scottmartinnet]

We've improved our approach to merging polygons from different tiles with a bitmap-based approach and an increased tile overlap percentage and added a filtering layer to ignore polygons at the boundary of a tile which do not continue onto other tiles.