search

eowjd0512 / fisheye-calib-adapter

Fisheye-Calib-Adapter: An Easy Tool for Fisheye Camera Model Conversion
https://www.linkedin.com/in/eowjd0512/
MIT License
71 stars 10 forks source link

choose of fisheye model #4

Open nieyan opened 3 months ago

nieyan commented 3 months ago

Thanks for your great work. I am confused of which fisheye model is better ? How to make the dicision? My camera's FOV is 195 x 150

Sheradil commented 3 months ago

FYI: I am not the author of this repository. Just randomly refreshes the issues because of my own issue.

I don't think that question can be answered right away without further information. What do you want to achieve?

The best camera model in terms of accuracy is the OcamCalib model. It can contain a higher number of parameters and is therefor able to capture more details.

However, if you want to use the resulting images for some AI processes that need the camera model parameters in advance, then you should know that almost every AI wants images that were taken by a pinhole camera. So it depends

eowjd0512 commented 3 months ago

Thank you for taking an interest in my project.

This project supports conversion between fisheye camera models, and it's difficult to definitively say which model is better as it can depend on the situation. As Sheradil mentioned, this system could be useful for converting when the necessary camera parameters are not directly available from the network model or system you want to use.

@Sheradil also noted that OCamCalib might be preferable for its accuracy, but you might not have access to a MATLAB to extract the parameters directly. Alternatively, since OpenCV supports the KB model, you might end up using the KB model depending on your calibration expertise (or some companies may only support the KB model for calibration). There are also datasets fixed to specific models. In situations where you're stuck with a fixed parameter model and need to convert it to another, this could be essential. For example, fisheye SLAM systems that take an EUCM or DS model as input, or object detection network models that only accept a UCM model.

In such cases, you could use this project to facilitate model conversion.

nieyan commented 3 months ago

FYI: I am not the author of this repository. Just randomly refreshes the issues because of my own issue.

I don't think that question can be answered right away without further information. What do you want to achieve?

The best camera model in terms of accuracy is the OcamCalib model. It can contain a higher number of parameters and is therefor able to capture more details.

However, if you want to use the resulting images for some AI processes that need the camera model parameters in advance, then you should know that almost every AI wants images that were taken by a pinhole camera. So it depends

I want to use the intrisic&extrisic of camera to get the depth of the pixel as accurate as possible; Especially at the edge of fisheye image where has larger distortsion.

Sheradil commented 2 months ago

I think with such a large field of view the OCam model is your best choice