This is an extension to YOLOv2, including our trained model for vehicle detection on the dataset of 2018 NVIDIA AI City Challenge.
We strongly encourage users to use the latest YOLOv4 object detector instead.
We select 4,500 frames uniformly sampled from the dataset of 2018 NVIDIA AI City Challenge, where each of them contains 5 to 40 objects. The training data are manually labeled in 8 categories, including sedan, hatchback, bus, pickup, minibus, van, truck and motorcycle. YOLOv2 is used for training and testing, where the pretrained weights are used to initialize the network.
- Download the trained model (weights) here.
- Follow the instruction in Joseph Redmon's blog to make and test the trained model. The given bash files provide examples of processing the entire dataset.
For output detection results in text, the format of each line is as follows:
<frame_id>,-1,<xmin>,<ymin>,<width>,<height>,<confidence>,-1,-1,-1,<class>
This is similar to the required format of MOTChallenge. The frame ID is 0-based. The confidence is in percentage.
For any question you can contact Zheng (Thomas) Tang.