PCA-RECT: Event-based Object Detection and Classification

The VLFeat Library is provided with the repo and the MATLAB script configures it on-the-fly.

Needs MATLAB AER Vision Functions from Garrick Orchard.

There are four versions of the code:

FPGAModular: For evaluating the exact modular version where FPGA is closely followed

FAST: 100x faster Quick floating-point versions for parameter testing (YH and RB)

FASTnoPCA: Fast version without principal component analysis (YH and RB)

FASTnoPCAwithDet: Fast version with detector incorporated (no PCA) (YH and RB)

The training files can be found in the N-SOD Dataset and needs to be placed in the correct path, relative to the main executing file. (the code uses '../' to reference the files).

../N-SOD Datatet/

Instructions to execute

1. Download the N-SOD Dataset and place above the PCA-RECT folder.
2. Add to path the MATLAB AER Vision Functions.
3. Run one of the scripts, e.g. Event_context_DEMOuav_rmax7by7rect_FAST

Instructions for Tunning Parameters (testing)

Tune descriptor size:

1. Set value of "param.descsize=7"
2. CTRL+H to replace "5by5" to "7by7"

Tune codebook size:

1. Set value of "histopts.num_bins=150"
2. CTRL+H to replace "100codebok" to "150codebok"

Then, you can run the code to load or get your data properly.

Citation

Ramesh B., Ussa A., Vedova L.D., Yang H., Orchard G. (2019) PCA-RECT: An Energy-Efficient Object Detection Approach for Event Cameras. In: Carneiro G., You S. (eds) Computer Vision – ACCV 2018 Workshops. ACCV 2018. Lecture Notes in Computer Science, vol 11367. Springer, Cham

@InProceedings{10.1007/978-3-030-21074-8_35,
   author="Ramesh, Bharath and Ussa, Andr{\'e}s and Vedova, Luca Della and Yang, Hong and Orchard, Garrick",
   editor="Carneiro, Gustavo and You, Shaodi",
   title="PCA-RECT: An Energy-Efficient Object Detection Approach for Event Cameras",
   booktitle="Computer Vision -- ACCV 2018 Workshops",
   year="2019",
   publisher="Springer International Publishing",
   address="Cham",
   pages="434--449",
   isbn="978-3-030-21074-8"
}