My project is based on object detection and the only class I am using is “Fire”. I have this rough idea about a drone project using deep learning and computer vision, and for that, I’ve used YOLOv5. I believe that a major part of the problem is because of the dataset that I am using. I have been recommended by my professors to use StyleGAN2 to synthesize fire dataset, but in my opinion, I think that this is completely irrelevant. Would be great if anyone could take a peek at my dataset.
By the way, hello @anon59033456, would you mind taking a look? The support key is jzZVpDuEzzeaP7r9IhHT2M5Bk5V2. I believe that this shoud provide access to my account. The problem that I think I am facing is with respect to the low precision and recall values. The loss and accuracy seems to be fine, at least to me, in both train and validation.
jzZVpDuEzzeaP7r9IhHT2M5Bk5V2 didn’t have a dataset, so I’m looking at o4PqLaQWeN2KZhtoPjOp which has 3.
I’m looking at Fire Annotation (V3) which has the most images. Here’s my feedback:
My idea is that our drone will use reinforcement learning, and computer vision to detect flame region,
^^ Grabbed that form your Reddit post. If you want your drone to detect fires, it would be beneficial for it to have images that were captured aerially vs what currently looks like images that were taken from a person on the ground. (Though I didn’t look through them all, this is what I mostly saw)
Question I like to ask myself before gathering data “How will this model be used in the wild?” Use that answer ti gather images of it will encounter in the wild.
I really like that you have null images in here. I would add aerial null images. Null aerial forest images, city, beaches, or anywhere the drone would be detecting in the wild.
Does this help? Happy to jump on a call as well 
Just saw your Reddit about being overseas. I’ll DM you a link to my Calendly
When you said your project isn’t doing well with certain metrics, which metrics were they?
I’m curious to see your metrics after you train on the Roboflow app. Looks like you have 10 free training credits. Would you be able to train the dataset you are looking to deploy a model for? Here’s how I would go about doing it with different versions to compare:
- v1 - train raw. no preprocessing or augmentation
- v2 - preprocessing: auto-orient, resize. Image level augmentation: Flip (horizontal and vertical), 90 degree rotate, crop, brightness, blur, noise. (you can train from your v1 checkpoint here)
I ask bc I feel like the current metrics are lower than I expected considering the current images you have and the amount of data. If you’d like like I can go ahead and train a version for you as well.
@Mohamed Would you be able to answer the rest of the questions Ashvith wants us to look into?
Sounds good. We’re here to help if you cross any blockers