I'm currently developping an application that uses BEV ( https://github.com/Arthur151/ROMP ) to detects humans on a Webcam video stream, and generate cropped images arround them. After that, I use a pretrained Reznet18 network to identify every person. But I have difficulties to make the identification working. So, I suppose it can be improved by doing some segmentation on the BEV's cropped results. I tried several classic segmentations methods : like threshold or gradiant segmentation, but it's not working. So, I'm searching a lib that can do that quickly, to do it during a live stream.
I'm not sure if PixelLib can be fast enough to do that. Perhaps it can by using a really small network ? Or...No network at all ? Is the segmentation part of the lib requiring a network ? Can I simply send the crops to it to get the segmentation areas without any type of identification, since I'm doing identification myself ? Or send the image with a an array containing the bounding boxes ?
Hi ! I have a question about PixelLib.
I'm currently developping an application that uses BEV ( https://github.com/Arthur151/ROMP ) to detects humans on a Webcam video stream, and generate cropped images arround them. After that, I use a pretrained Reznet18 network to identify every person. But I have difficulties to make the identification working. So, I suppose it can be improved by doing some segmentation on the BEV's cropped results. I tried several classic segmentations methods : like threshold or gradiant segmentation, but it's not working. So, I'm searching a lib that can do that quickly, to do it during a live stream.
I'm not sure if PixelLib can be fast enough to do that. Perhaps it can by using a really small network ? Or...No network at all ? Is the segmentation part of the lib requiring a network ? Can I simply send the crops to it to get the segmentation areas without any type of identification, since I'm doing identification myself ? Or send the image with a an array containing the bounding boxes ?
Thanks in advance for your answers !