Skip to content

Latest commit

 

History

History
51 lines (34 loc) · 1.66 KB

README.md

File metadata and controls

51 lines (34 loc) · 1.66 KB
Raw

Target Object Detection with SiamRPN+CenterNet

This project integrates the idea of the CenterNet into the SiamRPN universal object detection and tracking model without specific training. The SiamRPN uses features compared to find where is the template object in the search image. The CenterNet makes the detection model more simple to train and prepare data.

*This now still support for single object only


Demo

To use very large datasets to hopefully cover as many objects as possible. Surely it will use too much time to train too. So these demo models I provide are trained only for one epoch, Therefor it may still have low accuracy.

run main.py

    python main.py -v {video file} -m {path to .pth} -s {model slug (r18 or r50)}

Draw a bounding box on an object to be a target. If no --image args program will let you draw on video press 'c' to pass a frame, press 'space bar' to finish (you may have to press two time)

Data

ImageNet: ILSVRC

To use my ILSVRC loader the data folder format must be.

./dataset/ILSVRC/Data/VID/{train, val}

Training

later.

Reference