About the UI Demo

View the ITSD user interface demo here.

We have deployed a user interface demo using the Gradio framework. Currently, it supports both video and image detection via upload, with an adjustable confidence threshold. Several pre-uploaded examples are also provided. We have decided to use the YOLOv8 model (without bagging) for the demo due to its better performance and faster inference time (due to it being only one single model, and no aggregation is needed during inference).

The prediction results for photos are much better than videos. We take each frame of the video and process it to the model every 2 seconds. However, several aspects, such as object motion, camera movement, movement speed, change in lighting, etc. can make it harder for our model to detect traffic signs accurately. For example, fast-moving traffic signs may appear distorted or blurred, which may be harder to detect. Another problem is that traffic signs may be partially visible in one frame and fully visible in another. Other important conditions also exist, such as the weather, the capture time (day or night), et cetera.

Most of the traffic sign images in our data are clean, well-captured, and has good quality. Thus, the model may generalize better for static image input but perform worse with dynamic and noisy video input. For future improvement, we might need a more diverse dataset that includes images with motion blur, distortions, and varying lighting. It is also good to incorporate different kinds of driving environments (weather, capture time, etc.) into the dataset.