rozumden / FMODetect

[ICCV 2021] FMODetect: Robust Detection of Fast Moving Objects

FMODetect Evaluation and Training Code

FMODetect: Robust Detection of Fast Moving Objects

Denys Rozumnyi, Martin R. Oswald, Vittorio Ferrari, Jiri Matas, Marc Pollefeys

Inference

To detect fast moving objects in video:

python run.py --video example/falling_pen.avi

To detect fast moving objects in a single frame with the given background:

python run.py --im example/ex1_im.png --bgr example/ex1_bgr.png

We only provide the detection part. The deblurring and trajectory reconstruction part will be added later.

Dataset generation

Before generating the dataset, please make sure you cloned recursively, e.g. git clone --recursive [e-mail]:rozumden/FMODetect.git Also, please set your paths in dataset/generate_dataset.sh. Then, run this script.

Pre-trained models

The pre-trained FMODetect model as reported in the paper is available here: https://polybox.ethz.ch/index.php/s/X3J41G9DFuwQOeY.

Reference

If you use this repository, please cite the following publication ( https://arxiv.org/abs/2012.08216 ):

@inproceedings{fmodetect,
  author = {Denys Rozumnyi and Jiri Matas and Filip Sroubek and Marc Pollefeys and Martin R. Oswald},
  title = {FMODetect: Robust Detection of Fast Moving Objects},
  booktitle = {arxiv},
  address = {online},
  month = dec,
  year = {2020}
}