Recommend a model for depth estimation

General
, ,
1

Hello,

I would like to run object detection and depth estimation models to detect vehicles. I’ve picked YOLOv8 for object detection but confused between fast-depth or sc-depth for depth estimation.

My current hardware is, Raspberry pi 5 with Hailo 8(26TOPS) and a single Pi camera(open to adding ultrasonic sensors or stereo camera for accuracy) Please suggest a model that is flexible.

Also for future use can I get the output of the two models in json format?

Thank you

1 Like
2

Hey @varun.maale ,

Welcome to the Hailo Community!

For depth estimation with your Raspberry Pi 5, Hailo-8, and Pi camera setup, I can definitely help you evaluate the best model options! Based on your needs, here are the two main models I’d recommend considering:

FastDepth would be your best bet if you want something efficient and real-time. It’s specifically optimized for embedded systems like the Raspberry Pi and runs really well on the Hailo-8. The beauty of FastDepth is that it gives you solid depth estimation accuracy while still maintaining real-time performance, which is crucial for most applications. If you’re planning to stick with your single Pi camera setup, this would be my top recommendation.

However, if absolute accuracy is your priority, you might want to look at SC-Depth (Self-Calibrated Depth). While it’s a bit more demanding on resources, the Hailo-8’s 26 TOPS should handle it well. The self-calibrating feature is particularly impressive for getting precise depth measurements. This would be especially valuable if you’re thinking about upgrading to a stereo camera setup in the future.

Regarding your sensor setup, you’ve got several options:

  • Your current single Pi camera can work well, especially with proper calibration
  • A stereo camera setup would significantly improve accuracy if you need it
  • You could even add ultrasonic sensors for specific distance checks, though they’re more complementary than essential

For the output format, I can help you structure everything in JSON. Here’s what I’m thinking:

  • YOLOv8 will handle object detection, giving you classes, confidence scores, and bounding boxes
  • The depth model (whether FastDepth or SC-Depth) will provide depth maps that we can convert into a more manageable JSON format
  • We can combine both outputs into a single, clean JSON structure

The final choice really depends on your specific needs. If you’re focusing on real-time processing with your current single camera setup, I’d recommend FastDepth. But if precision is your priority and you might add more sensors later, SC-Depth would be the way to go.

Would you like me to elaborate on any of these aspects or help you think through which option would best suit your specific use case?

1 Like
3

Thank you for the detailed reply,

here are more details about the use:
The setup would be fitted on a vehicle with a max speed of 7 meters per second (25 Kph) and ideally I would like the setup to detect other road users 70 meters (10 seconds) ahead so that the detected data from the live camera feed is sent to other processor and has sufficient time to react.

I’m leaning towards sc-depth along with sensors as improved accuracy results in better safety. But I would like your input as well based on the above use case.

Thank you.

4

Let me clarify why I think this setup would work well for your vehicle detection system.

Looking at your 70-meter detection range requirement, SC-Depth would be your best option. It’s particularly good at accurate depth estimation at longer distances, which is exactly what you need.

While a single Pi camera can get the job done, you might want to consider either:

  • Adding a stereo camera setup to improve depth accuracy, especially when you’re moving at higher speeds
  • Including ultrasonic sensors as a backup for close-range detection

For combining YOLOv8’s detection with SC-Depth’s measurements, we can output everything in a clean JSON format that’s easy for other systems to process. This would make it simple to send the data to other processors in your vehicle.

Can Wait to see this project working,
Best Regards