Skip to content

Latest commit

 

History

History
100 lines (78 loc) · 4.09 KB

README.MD

File metadata and controls

100 lines (78 loc) · 4.09 KB


Paper | Blog Post
This repository will enable you to quickly perform visualization, inference, evaluation and training on our Expo Markers Dataset with Mask R-CNN / Faster R-CNN using Detectron2.
See our Colab Notebook.

Visual Results of our trained netweork on our real image test set.

The EXPO Markers Dataset

The EXPO Markers Dataset is a dataset of Expo whiteboard markers for the purpose of instance segmentation and object detection. The dataset contains three subsets (all include instances segmentation labels):

  • Expo_Synt_V8 Photorealistic synthetic image dataset with 5000 images (1024x1024).
  • Expo_Real_DGOffice Real image dataset with 250 images (used for validation and test).
  • Expo_Real_India Real image dataset with 1000 images (used for training and comparison to our synthetic data).

Our Code

About detectron2

Our code is using detectron2 and is affected by detectron2's examples.

https://github.com/facebookresearch/detectron2

Read their blog post to see more demos and learn about detectron2.

Installation

Install detctron2 using their instructions.

Download the Expo Markers Dataset and Pretrained Weights

The data can be downloaded from our blog post.

Quick Start

in detectron2, both the trainer and the predictor objects, are initialized using a config object which in our code is constructed from three sources with the following hierarchy: args.config_file < datagen_config < args.opts.

  1. args.config_file - optional (default='').
  2. datagen_config - a json file located at settings/datagen_config.json with our default parameters.
  3. args.opts - parameter settings from argparse. see settings/datagen_setup.py.

This Colab Notebook will show you how to use our code.

Inference with Pre-Trained Models

python inference/inference_markers.py

Evaluation with Pre-Trained Models

python evaluation/eval_markers.py \
MODEL.WEIGHTS /path/to/weights/model.pth \
DG_TEST_SET_PATH /expo_markers/expo_datasets/real_image_dataset/ \
DG_TEST_SET_INDS "('0-50',)" \
DG_EVALUATION_OUTPUT_PATH /output/path/eval.pickle

Training

python training/train_markers.py \
DG_TRAIN_SET_PATHS "('/path/to/dataset/dir/',)" \
DG_TRAIN_SET_INDS "('0-750',)" \
DG_TRAINING_OUTPUT_DIR /path/to/output/weights/dir/ \
SOLVER.MAX_ITER 30000 \
SOLVER.STEPS '(25000, 27000)'

License

The Datagen expo dataset is released under Non-Commercial Data License Agreement.

Citing

If you use our dataset or code in your research please use the following BibTeX entry.

@misc{ron2021detection,
      title={Detection and Segmentation of Custom Objects using High Distraction Photorealistic Synthetic Data}, 
      author={Roey Ron and Gil Elbaz},
      year={2021},
      eprint={2007.14354},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}