wanjinchang / yolo-lite

All the trained models used while developing YOLO-LITE

YOLO-LITE

YOLO-LITE is a web implementation of YOLOv2-tiny trained on MS COCO 2014 and PASCAL VOC 2007 + 2012. All the trained models (cfg and weights files) used while developing YOLO-LITE are here. Our goal is to create an architecture that can do real-time object detection at a speed of 10 FPS and a mean average precision of about 30% on a computer without a GPU.

Demo

Check out our models trained on COCO and VOC here.

Below is the COCO YOLO-LITE model performing real-time object detecion at about 10 FPS from a Dell XPS 13 laptop:

Results

DataSet	mAP	FPS
PASCAL VOC	33.57	21
COCO	12.26	21

best PASCAL cfg | best PASCAL weights

best COCO cfg | best COCO weights

*Note: FPS is calculated from runnig locally on a Dell XPS 13 laptop.

Get Started

Training

We used AlexeyAB's Darknet for Windows to train our model. Install Darknet here.

To find the mAP for each training model, run the command under the scripts folder:

python mapScript.py

When prompted, add the location of the cfg and the location of the weights folder.

Testing

In order to get the FPS, we used a Python adaptation of Darknet called Darkflow here.

Web Implementation

1. To convert the model to JavaScript, we followed the following tutorial.

2. Once converted to the JavaScript, refer to our two repositories of tfjs-yolo-tiny and tfjs-yolo-tiny-demo.

3. Replace line 14 in index_coco.js and index_voc.js to a link from the resulting .json file in Step 2:

   model = await downloadModel('put your link here');

How to prune weights

While we found that pruning weights by a simple threshold did not really effect the mAP or FPS somoene else may find this script useful. To prune a weights file navigate to the weights.py file then run

python weights.py WEIGHTS_FILE_NAME THRESHOLD