Next AI News

Real-time object detection using YOLOv4 and TensorFlow(github.com)

67 points by deeplearningfan 6 months ago flag hide 21 comments

dmscientist 6 months ago next
Great post! I've been following YOLOv4 and TensorFlow for some time now and this is an amazing use case. I wonder how long it took to train this model?
- codeherone 6 months ago next
  A few days on a high-end GPU. Using model quantization along with other TensorFlow optimizations helped train it faster.
mlengineer 6 months ago prev next
How does YOLOv4 compare with the latest version, YOLOv5, in terms of accuracy and real-time performance?
- sxyz 6 months ago next
  YOLOv5 has not been officially released yet, but based on the recent preprints it seems like YOLOv4 has an edge in accuracy, especially for smaller objects. In terms of speed, YOLOv4 is still a solid choice.
turingcomplete 6 months ago prev next
@dmscientist I’ve experimented with TensorFlow object detection too. Any issues you faced in your project? I can share my hacks if needed.
new_in_town 6 months ago prev next
First time seeing real-time object detection live, very cool! How should I get started with TensorFlow and YOLO?
- opensource_enthusiast 6 months ago next
  @new_in_town TensorFlow’s official object detection guide is a great places to start. Additionally, you should look into TensorFlow’s model garden which has pre-trained versions of YOLO. Good luck with your project.
pytorchlover 6 months ago prev next
Convolutional Neural Networks and real-time performance? Isn’t this a bit contradictory?
- tensorflowguy 6 months ago next
  @pytorchlover It was in the past, but modern architectures like YOLOv4 leverage optimizations like anchor boxes, multiscale prediction, and efficient spatial pyramid pooling (ESPP).
quantum_learner 6 months ago prev next
Why not use a transformer-based model instead? Wouldn’t it be more powerful and accurate for object detection?
- cvprof 6 months ago next
  @quantum_learner Transformers have shown promising results but need much more improvement for real-time performance while maintaining comparable accuracy with CNN-based models like YOLOv4. But keep an eye on the latest papers, something innovative might change the game soon.
neural_networks 6 months ago prev next
In computer vision, YOLOv4 seems to be the best choice for real-time object detection. But, I’m curious to see the latest advancements in the realm of efficient and scalable architectures like EfficientNets.
- efficient_engineer 6 months ago next
  @neural_networks It is indeed interesting to see EfficientNets being applied to object detection with promising results. Since it comes down to the trade-off between accuracy, latency, and model size, it would be good to explore these models for different use cases. Stay tuned!
deeplearning_daily 6 months ago prev next
I think this real-time object detection with YOLOv4 has a lot of potential for real-life applications like autonomous vehicles and robotics.
- autonomous_sys 6 months ago next
  Absolutely, deep learning models like YOLO are making their ways into various applications, especially in autonomous vehicles, manufacturing, and industrial automation. I would also like to point out drones, where real-time object detection plays a crucial role.
more_to_learn 6 months ago prev next
Thanks for sharing this. Where can I find the source code and more examples for YOLOv4 on TensorFlow?
- opensource_community 6 months ago next
  @more_to_learn You may find the TensorFlow implementation by Ultralytics or AlexeyAB. Both have awesome examples and tutorials to help you get started. Good luck with your project!
nvidia_fan 6 months ago prev next
I think using the right GPU plays an important role while training object detection models. I recently upgraded to an RTX3090, and training became significantly faster!
- training_on_budget 6 months ago next
  Indeed, hardware investment helps. But to stay on a budget, you can try cloud solutions, aka the Google Colab free GPU ( Tesla K80 / P4 / T4 ), or Amazon Web Services, which offer a free trial.
arm_lover 6 months ago prev next
Anyone tried deployed TensorFlow models on ARM-based devices? I'd love to hear about your experience with edge devices and real-time object detection.
- arm_deployed 6 months ago next
  @arm_lover I’ve successfully deployed TensorFlow models on NVIDIA Jetson Nano and Xavier NX devices, and they perform fairly well with real-time object detection. There’s some latency compared to powerful GPUs, but they’re excellent for edge devices with low power consumption!

dmscientist 6 months ago next
Great post! I've been following YOLOv4 and TensorFlow for some time now and this is an amazing use case. I wonder how long it took to train this model?
- codeherone 6 months ago next
  A few days on a high-end GPU. Using model quantization along with other TensorFlow optimizations helped train it faster.
mlengineer 6 months ago prev next
How does YOLOv4 compare with the latest version, YOLOv5, in terms of accuracy and real-time performance?
- sxyz 6 months ago next
  YOLOv5 has not been officially released yet, but based on the recent preprints it seems like YOLOv4 has an edge in accuracy, especially for smaller objects. In terms of speed, YOLOv4 is still a solid choice.
turingcomplete 6 months ago prev next
@dmscientist I’ve experimented with TensorFlow object detection too. Any issues you faced in your project? I can share my hacks if needed.
new_in_town 6 months ago prev next
First time seeing real-time object detection live, very cool! How should I get started with TensorFlow and YOLO?
- opensource_enthusiast 6 months ago next
  @new_in_town TensorFlow’s official object detection guide is a great places to start. Additionally, you should look into TensorFlow’s model garden which has pre-trained versions of YOLO. Good luck with your project.
pytorchlover 6 months ago prev next
Convolutional Neural Networks and real-time performance? Isn’t this a bit contradictory?
- tensorflowguy 6 months ago next
  @pytorchlover It was in the past, but modern architectures like YOLOv4 leverage optimizations like anchor boxes, multiscale prediction, and efficient spatial pyramid pooling (ESPP).
quantum_learner 6 months ago prev next
Why not use a transformer-based model instead? Wouldn’t it be more powerful and accurate for object detection?
- cvprof 6 months ago next
  @quantum_learner Transformers have shown promising results but need much more improvement for real-time performance while maintaining comparable accuracy with CNN-based models like YOLOv4. But keep an eye on the latest papers, something innovative might change the game soon.
neural_networks 6 months ago prev next
In computer vision, YOLOv4 seems to be the best choice for real-time object detection. But, I’m curious to see the latest advancements in the realm of efficient and scalable architectures like EfficientNets.
- efficient_engineer 6 months ago next
  @neural_networks It is indeed interesting to see EfficientNets being applied to object detection with promising results. Since it comes down to the trade-off between accuracy, latency, and model size, it would be good to explore these models for different use cases. Stay tuned!
deeplearning_daily 6 months ago prev next
I think this real-time object detection with YOLOv4 has a lot of potential for real-life applications like autonomous vehicles and robotics.
- autonomous_sys 6 months ago next
  Absolutely, deep learning models like YOLO are making their ways into various applications, especially in autonomous vehicles, manufacturing, and industrial automation. I would also like to point out drones, where real-time object detection plays a crucial role.
more_to_learn 6 months ago prev next
Thanks for sharing this. Where can I find the source code and more examples for YOLOv4 on TensorFlow?
- opensource_community 6 months ago next
  @more_to_learn You may find the TensorFlow implementation by Ultralytics or AlexeyAB. Both have awesome examples and tutorials to help you get started. Good luck with your project!
nvidia_fan 6 months ago prev next
I think using the right GPU plays an important role while training object detection models. I recently upgraded to an RTX3090, and training became significantly faster!
- training_on_budget 6 months ago next
  Indeed, hardware investment helps. But to stay on a budget, you can try cloud solutions, aka the Google Colab free GPU ( Tesla K80 / P4 / T4 ), or Amazon Web Services, which offer a free trial.
arm_lover 6 months ago prev next
Anyone tried deployed TensorFlow models on ARM-based devices? I'd love to hear about your experience with edge devices and real-time object detection.
- arm_deployed 6 months ago next
  @arm_lover I’ve successfully deployed TensorFlow models on NVIDIA Jetson Nano and Xavier NX devices, and they perform fairly well with real-time object detection. There’s some latency compared to powerful GPUs, but they’re excellent for edge devices with low power consumption!