Implemented with TensorFlow Object Detection API.
Tested on LaRA dataset.
Model inference example:
Check out the rendered video in Youtube or BaiduPan
On-board vehicle acquisition in a dense urban environment:
- 11179 frames (8min 49sec, @25FPS)
- 640×480 (RGB, 8bits)
- Paris (France)
Links:
-
Download Dataset download link
-
Download Ground truth labels
To make TFRecord files for Tensorflow tranning, read this
Here records an informal test performance on 592 unseen images:
- Model = SSD MobileNet, pre-trained on COCO
- Infer time per image = 9 ms
- Green light [email protected] = 0.385
- Red light [email protected] = 0.725
- Yellow light [email protected] = 0.385
- Precision [email protected] = 0.620
Running on Tesla P40 GPU
Training total loss:
Do git clone https://github.com/tensorflow/models.git and update directory in .sh files
Follow the instructions at this page for installing some simple dependencies.
Location of pre-trained models: pre-trained models zoo
Download the required model tar.gz files and untar them into models/ directory with tar -xvzf name_of_tar_file.
python data_conversion.py --input_yaml lara/annotations_train.yaml --output_path lara/train.record
python data_conversion.py --input_yaml lara/annotations_test.yaml --output_path lara/test.recordsh train.sh <faster_rcnn | ssd_inception | ssd_mobilenet>
sh evaluate.sh <faster_rcnn | ssd_inception | ssd_mobilenet>
tensorboard --logdir=models --port=8052note you'd better not run train & evaluate together because they will use up GPU memory
sh freeze.sh <faster_rcnn | ssd_inception | ssd_mobilenet> <model checkpoint version num>using the TrafficLightDetection-Inference.ipynb