Drone Model, high overlap. #548
Replies: 3 comments 1 reply
-
|
Can you show some data and model training that fails given the current retinanet backbone? We have always believed that data, not architecture, is the limiting factor here. No doubt that the new YOLO models have great functionality and speed, but given the data sizes the vast majority of ecological projects, I have yet to see an example in which the two backbones, given the same data, produce meaningfully different, let's say 4% results. I am 100% open to including new models, and we hope to make the framework open to any particular kind of model a user wants to bring in. Having examples in which the current workflow is insufficient, but another model is better, would help immensely in speeding up this work. As an aside, we are retraining the detection backbone currently, are you able to share annotations and or images so that it can be improved for your use case? See #340 |
Beta Was this translation helpful? Give feedback.
-
|
Hello, thank you for your prompt reply. I'm encountering difficulties with the following types of images:
I'm using a total of 169 images for training. To address the issue of high overlap, I've increased the nms_thresh parameter, which resulted in the following metrics: Box Precision: 0.9 Thanks |
Beta Was this translation helpful? Give feedback.
-
|
Great, I've made this into a discussion. Can you drop any code you've used and I can look through it. If you can share the data I will have a quick run and see if I can recreate your score. I'm sure we can do better than that recall, I'll know more looking at the code, but that precision score is quite high compared to recall. The images look preprocessed in some interesting way, how were they acquired? Can you tell me about the project? |
Beta Was this translation helpful? Give feedback.
-
|
Hi, thanks for the response, My code is in this repo, and I am running it using the next colab. Unfortunately, I cannot share with you the raster itself (even though I have shared the png images in the colab), but I can give you some metadata about it:
About the preprocessing, I do the following preprocessing. |
Beta Was this translation helpful? Give feedback.
-
Please describe a feature you would like to be added.
I find it challenging to accurately detect and classify different species of trees in various landscapes using current image recognition software.
Describe the solution you'd like
I would like an integration with YOLOv8 from Ultralytics, which is state-of-the-art in object detection. This could enhance the ability to discern between tree species by leveraging YOLOv8's high precision and real-time processing capabilities.
Beta Was this translation helpful? Give feedback.
All reactions