Auto-Lama

Auto-Lama combines object detection and image inpainting to automate object removals. It is build on top of DE:TR from Facebook Research and Lama from Samsung Research. The entire process is extremely simple:

  1. Objects are detected using the detector.
  2. Masks are generated based on the bounding boxes drawn by the detector.
  3. The original image is sent to the inpainter along with the masks.

Demo

cat
buildings
park

Masking

There are currently a few ways of generating masks:

  1. Masking objects with specified indices.
  2. Masking one main object at a time.
  3. Masking all other objects other than the main object.

Future Goals

  1. Use a more precise segmentation method other than bounding boxes
  2. Implementing a detector that has more

Environment Setup

Prerequisites

  1. docker
  2. make
  3. conda

Building Environment

  1. make build-conda-env
  2. conda activate auto-lama
  3. make build-env

Cleaning Directory

  1. make clean

Detect and Inpaint

Setup

The default config for the detector is

PARAMETERS = {
    "model_name": "facebook/detr-resnet-50",
    "threshold": 0.9,
    "max_items": 10,
    "save_destination": "./test_images",
    "output_destination": "./output_images",
    "max_width": 2000,
    "max_height": 2000,
    "resize": True,
    "resize_scale": 0.75,
    "excluded_objects": [91],
    "image_format": "PNG",
    "mask_target_items": [],
}

Please reference here for the target items that you want to mask, as the default DE:TR uses the COCO Dataset,

Run

  1. make detect_and_inpaint IMAGE_PATH=path/to/image or make detect_and_inpaint IMAGE_PATH={image_url}

GitHub

View Github