Should a webcam configuration file be changed to use to get a better match with the newly trained models?
For example, we are using a Logitech C920 webcam which can get high resolution, but the trained model is 320x320. Would it be better to adjust the webcam to start it at a lower resolution? The document discusses this, but should something be done in the webcam config file to address this?
At face value it appears that there would be SOME benefit in configuring the camera using a webcam configuration file. The closer you can get to a “square” image the fewer stretching effects will be created in the training data. Also I would assume the camera hardware would do a better job downscaling the image to the configured resolution, and the smaller the “starting” image the fewer downscaling effects you’ll get as well as the image is closer to the target model resolution. Finally, it seems like there would be less processor utilization in reducing the images, which may improve processing speed on both the training and on-robot detecting speed.
With that said, I have never tried to characterize the difference in quality or performance of processing the images. The lion’s share of the performance hit is the actual operation of the inference model, and not necessarily in the processing of the individual image snapped from Vuforia. The detection quality will likely be different, but my gut says it won’t be crazy different simply because of the neural network that is already trained to be generalized - there will be a difference, but there’s no clear indication that it’s “significant”. My gut says it might be the difference between a couple percentage points in detection quality, depending on the kinds of objects you’re trying to detect/categorize, which in itself wouldn’t be all that significant, but again I don’t know for sure.
The important element is that if you use a web configuration file, you really want to make sure you always use that web configuration file (during training and during on-bot usage).
If you want to do this as an additional test, send me an email and let’s talk about test parameters. I could justify some extra training time for your team to do this for the betterment of everyone.