I’m running the pose estimation from the hailo-rpi5-examples with video acquired from the theta x 360 degree camera but it doesn’t recognize people further than 3 meters from the camera. How can I improve to recognize people up to 10 meters away?
Welcome to the Hailo Community!
To ensure accurate predictions, it’s essential to consider the original dataset and the input size of the network. Pre-processing involves transforming the original image to fit the network’s input size, which may include resizing, cropping, and adjusting the image dimensions. The resulting image is the data the model was trained on.
Ensure that images captured by your camera undergo appropriate pre-processing so the input fed into the model closely resembles the training data.
For instance, if your camera has a wider field of view, objects or people farther away will appear smaller. When these images are resized and cropped, there might not be enough pixel detail for the model to infer the correct results accurately. In such cases, you may need to adjust the tiling and resizing strategy of the sensor’s images.
If the pre-processing transformations result in slightly distorted or squeezed images, consider retraining the model using this altered data. Doing so can improve the model’s performance on your specific image inputs.