(QUESTION) Which model to use for multipas detections

[MatheusCaltran]

Hello JK Jung

With your repositories I can learn more about AI, so I have a lot of consideration for you

I would like to know in your opinion which model I can use to detect several objects?(as shown in the image I have to do the cane detection in an agricultural machine)

I'm thinking of using something like yolov4 tiny and leave the jetson nano inside the machine, or to use the yolov4 with a jetson xavier(I've done some training using SSD but I didn't have a good detection) Thanks in advance.

[jkjung-avt]

In general, yolov4 is a good choice for object detection applications.

For your specific application, I think how you label the data and train the model would be crucial.

[MatheusCaltran]

thank you I hope one day to have a little of your knowledge, anything I repost the topic to ask about something else. But I would try to avoid it so as not to occupy your time. For now I will retrain the SSD with more data, then I will try with yolov4tiny using your example Demo #5: YOLOv4(I only have jetson nano and google colab available)

[agjunyent]

If you only need to detect the canes (1 class), in this environment, yolov4-tiny-3l from darknet (https://github.com/AlexeyAB/darknet) will be more than sufficient, with good fps. If you increase the number of classes but maintain the "aspect ratio" of them (so all classes have similar sized boxes), yolov4-tiny-3l still is a very good candidate.

Crucial parts:

• Labelling, as @jkjung-avt said. I'd say that 50% of the success of a model, or even more, is training dataset labels quality
• Number of training samples. The more the better, but slower to label
• mAP requierements. You need to think that you will never have 100% accuracy / mAP.

[MatheusCaltran]

thank you very much, i will try with yolov4-tiny-3l sorry to take up your time

[MatheusCaltran]

I liked the result but I think it can be improved, I'm going to add more data to my AI I'm also thinking about testing YOLOV7, but I would like to know if there's any camera you recommend to use in this case (in jetson nano and and with good quality) currently I'm using C270 and I'll try to use a CSI e-cam30_cunano. I'm not sure which is better CSI or USB?

[jkjung-avt]

@MatheusCaltran I think the choice of camera depends on your application. In general,

USB webcam:

• Pros: cheaper and easier to get; usually doesn't require additional device driver, so easier to use.
• Cons: resolution & frame rate are limited by USB bandwidth (usually lower than CSI cameras); consumes more CPU for grabbing video data.

CSI camera:

• Pros: higher resolution & frame rate; lower latency; has the option of adjusting ISP settings to achieve better video quality or special effects.
• Cons: more expensive; usually requires additional device drivers; can only run very short cables.

[MatheusCaltran]

thanks I will focus on using CSI, I will use ECAM-30 for now

[MatheusCaltran]

Hi @kjung-avt how are you?

we have already made several modifications to the machine, I am currently using jetson nano, e-cam30, yolov4-tiny

detection photo:

but now we do a test simulating reality and it didn't work out very well because there was straw, but this will only happen in extreme cases because of the rain that makes the straw stick to the sugarcane

example:

With all this I just want to know about your opinion what can I do to improve this? for me your opinion is very important, much of what I can do here was because of you.

what I think about doing is training with the straws, but I'm sure the detection won't be very good even with the training.

[jkjung-avt]

Adding pictures with straws into your training set would definitely help. I think you should do that.

I think a trained model should be able to detect sugarcanes which are partially covered by straws and still clearly visible.