Details:
• Hardware - Tesla P4
• Network Type - DINO
• Running TAO models in TensorRT
• Deployable weights
I’m having problems with the retail object model not detecting many small-ish items. My goal is to be able to detect and track retail items throughout a video scene accurately, and I am having problems with just identifying typical items. I have tried using the various model types available on the NGC model zoo (V2.2.2.1, V2.2.2.2, V2.2.2.3, etc) and each performs slightly differently, but without a clear winner.
The image below shows two items that are much larger than the minimum 10x10px size the detector is claimed to catch. The are outlined in the dotted red boxes. Neither of these items is ever recognized in a 30 second scene of a 30fps video.
Questions
- Are there methods to increase the performance of these models?
- Is there any preprocessing of images required for these models before inference?
- What is the particular application of these models? Are they appropriate for retail settings?
3 posts - 2 participants