AI Community Models

Model Technical Details

PERFORMANCE METRICS

88.4% Average Precision

This model achieves an Average Precision score of 0.8843 for vehicle detection on the test subset of the combined WASABIand VisDrone datasets. This model performs best on imagery at the scale and resolution typically utilized by FMV sensors. For fast inference time, this model works best on a GPU.

A higher precision score indicates that on average the majority of labels predicted by the model for different classes are accurate.

OVERVIEW

The model utilizes the YOLOv3 architecture, a single-shot object detection architecture, only requiring a single forward pass of the input through the network in order to make detection predictions. In order to detect objects of different sizes, YOLOv3 detects objects at 3 different scales at different branches throughout its architecture. The prediction branch at each scale is responsible for predicting candidate bounding boxes and object classes. These candidate bounding boxes are then consolidated using a technique known as Non-Maximum Suppression to generate the final output.

TRAINING

This model was trained on a combination of the open source VisDrone dataset, as well as the WASABI dataset with custom vehicle labels. Any non-vehicle labels were removed. The following VisDrone labels were combined into one object class of “vehicle” — “car”, “van”, “bus”, and”truck”. The model was trained for 90 epochs.

VALIDATION

The model was evaluated on a test subset of the combined VisDrone and modified WASABI datasets.

INPUT SPECIFICATION

The input(s) to this model must adhere to the following specifications:

OUTPUT DETAILS

This model will output the following:

The output file (results.json) will contain detected vehicle bounding boxes. Each bounding box will contain the corresponding object score, class score, and top left/bottom right x,y coordinates defining the box.