QuickYOLO

Demo Video (Click on the image below)

Description

This repository is an end-to-end example for training a binarized object detector on the PASCAL VOC dataset, exporting it for inference, and using the exported model in a demo application.

It is forked from https://github.com/pythonlessons/TensorFlow-2.x-YOLOv3, which had the most comprehensive training pipeline based on TensorFlow 2 that I can find for training a YOLO detector.

Libraries used:

Larq 0.11.2
TensorFlow 2.3.2 (Unexpected behaviour may show up if using TensorFlow 2.4, see this issue)
OpenCV 4

Platform tested on:

NVIDIA Jetson Nano 2GB

The example model is based off YOLOv2, with QuickNet as the backbone, and a head that is based off the transition blocks in QuickNet. Here are the statistics for this model:

mAP@0.5 for VOC2007 Test Set: 0.37
Average Inference Time (Tested using the LCE benchmarking tool on a Jetson Nano): 26.5 ms

Demo through your computer (no speedups, just a rather meh detector)

cd demo_application
python demo_lce.py

End-to-End Usage

For a quick demo, just skip to the deployment step with the provided model.tflite and labelmap.tflite in demo_application/

1. Prepare dataset

Run dataset_preparation/dataset_preparation.sh / dataset_preparation/dataset_preparation_no_download.sh if already downloaded dataset
Run python dataset_preparation/migrate_data.py
Run python dataset_preparation/flatten_voc.py
Run mkdir voc_data
Run mv VOCdevkit/train voc_data && mv VOCdevkit/test voc_data
Run python dataset_preparation/remove_train_test_duplicates.py
(Optional) If you want to make sure the datset prepared is valid, run python dataset_preparation/check_voc_dataset_sanity.py
Run python tools/XML_to_YOLOv3.py

(Cleanup in case of messups)

Run rm -rf VOCdevkit/ && rm -rf VOC2007Trainval/ && rm -rf voc_data
Repeat the process with dataset_preparation_no_download.sh in step 1 instead.

2. Train

Run python train.py

3. (Optional) Evaluate accuracy

Change YOLO_CUSTOM_WEIGHTS in configs.py to the checkpoint for the trained model, e.g. checkpoints/quickyolov2_custom
Run python evaluate_mAP.py

4. Export

If haven't, change YOLO_CUSTOM_WEIGHTS in configs.py to the checkpoint for the trained model, e.g. checkpoints/quickyolov2_custom
Run python export_larq_model.py

5. Deploy

Install OpenCV4 (run demo_application/install_lce.sh from current directory)
Install Larq Compute Engine (run install_lce.sh from current directory)
Move demo_application/detection.cc to ${larq-compute-engine}/examples
Move demo_application/Makefile to ${larq-compute-engine}/larq_compute_engine/tflite/build_make/Makefile
Build using larq_compute_engine/tflite/build_make/build_lce.sh --native
Move the exported model, and voc_names.txt to gen/linux_aarch64, renaming them to model.tflite and labelmap.txt respectively.

References

A lot of the training code was forked from https://github.com/pythonlessons/TensorFlow-2.x-YOLOv3
Demo application took reference from https://github.com/finnickniu/tensorflow_object_detection_tflite/blob/master/demo.cpp

TODOs

Clean up the code base in this repo
Incorporate spatial pyramids
Make example application buildable using Bazel
Add multiprocessing for camera I/O in demo application

License

MIT

tehtea/QuickYOLO