Should I buy the shorter shutter speed industry camera to recognize the fast moving anmial?

I am going to use multi cameras taking 3D videos of an running monkey. I want to recognize the monkey body movement including toes. The toes will blur while the monkey is running. Will the blur toes image effect the recognition rate?

Should I buy the industry camera because it has the shorter shutter speed? What is the image max input size (resolution) of DeepLabCut?

Blurry frames may work worse both for labelling precision and models detections. Default max size is 1500x1500 px, but you can change it if you want to. Machine vision cameras provide a range of benefits, not only shutter speed control so, imo, if you can afford it you should definitely buy them. Especially if using them for 3D reconstruction (since you can have an external trigger and near perfect synchronization - in microsecond range - without using sound to sync)

1 Like

I am currently running a camera at 15 FPS 1080p, a camera at (supposedly) 50 FPS and a raspberry pi camera at 80 or 110 FPS to monitor rats.

DLC struggles with the 15 FPS camera due to the way the animal blurs through the frame during its maximum speed, but it seems to work fairly well with the (supposedly) 50 FPS or higher framerates, provided I give it enough illumination. Sometimes the blur is so bad I can’t even tell exactly what’s going on, so I abandoned the 15 fps cameras for DLC.

The other two cameras I use seem to work pretty well, apart from some other issues I’m trying to debug on this forum with people.

I have also tried DLC on data from two other collaborators, and from all 3 projects, I can tell you that sufficient contrast and lighting is also important. Blurring depends on what you want to do with the data and how bad the blurring is.

In terms of contrast, DLC did surprisingly well with a brown mouse on a brown background, but definitely had confusion some of the time. (Now the background is painted white, and I am awaiting video files.)

In my project, adding illumination to make the scene reasonably lit substantially boosted DLC performance, mainly because it added enough contrast and object detail.

In the third project, DLC was able to learn what a rat’s foot looked like, even though the foot was blurry in the video half the time. It would have been impossible to discern details of the foot shape with the blurring, but that wasn’t a requirement for this project. We just wanted to know the position of various joints in the hip/leg structure and didn’t care about the configuration of the foot, as long as we could identify its location.

A Raspberry Pi might give a relatively cheap way to test things out, although the pixel resolution goes down with increased temporal resolution on those. There are programmatic ways to control the camera acquisition with Python, if that is of interest.

tl;dr; if you want to be able to track the joints of the feet/hands/fingers, you will probably want relatively clear images.

1 Like

Hi, thanks for your reply. As far as I know, the shutter speed of the camera controls weather the image is blurring or not. If you want to take a clear image, you need to use a short shutter speed camera.
Thanks for telling me to add illumination.
I hope you can have a nice experiment result.

While I have no experience with DLC specifically, another aspect when selecting cameras for fast motion is whether the camera has rolling shutter or global shutter. For fast motion, rolling shutter can cause the imaged object to look deformed.

1 Like