-
Project Type:Solar Panel Research-using object detection images to train and detect Solar Panel defect of Bypass_Diode, Thermal_Hotspot, Cell_Fault. Using Classification and YOLO to compare results. understanding annotations and comparing both models
-
Operating System & Browser: Using VS Code to clean data and Google Colab to Train AI models
-
Project Universe Link or Workspace:henrys-workspace-tv1in/Project ID:solar-panel- thermal- detection
-
Do you grant Roboflow Support permission to access your Workspace for troubleshooting? (Yes/No):Yes
I have a thermal solar-panel dataset collected from Roboflow Universe. The images contain three defect classes: Bypass_Diode, Thermal_Hotspot, and Cell_Fault. The datasets were originally created for Object Detection, so the images have bounding-box annotations.
My research goal is to use these images to compare YOLO Object Detection vs Multi-Label Image Classification while keeping the experiment as fair as possible.
I am trying to understand the correct workflow for preparing the same Object Detection dataset for both types of models.
1. YOLO: Can I use the existing images and bounding-box annotations directly to train YOLO models?
2. Multi-Label Classification: How should I convert the same Object Detection images into full-image multi-label classification without cropping the bounding boxes? For example, if an image has 3 Cell_Fault boxes and 1 Thermal_Hotspot box, I want the classification label for the whole image to be:
Bypass_Diode = 0, Cell_Fault = 1, Thermal_Hotspot = 1
3. Same dataset: What is the best way to make sure YOLO and the classification models use the same source images and the same Train/Validation/Test split, so the comparison is fair?
4. Composite images: Some images contain multiple different defects. Besides full-image multi-label classification, what other methods should I consider for handling these composite images? For example, should I consider bounding-box cropping + classification, YOLO alone, or a YOLO + classifier pipeline?
5. Training outside Roboflow: I plan to clean the dataset in VS Code and train models in Google Colab. What Roboflow export formats/workflow would you recommend for maintaining the annotations and labels correctly for both experiments?
6. What exactly am I seeing when I view an annotated image in Roboflow?
When I open an image in Roboflow, I see the original thermal image with colored bounding boxes and class names such as Cell_Fault, Thermal_Hotspot, or Bypass_Diode.
Are the boxes and class names I am seeing the ground-truth annotations created by the person who labeled the dataset, or are they already an output/prediction from an AI model?
I am trying to understand how YOLO actually “sees” this during training. Does the model receive:
Original image pixels + class ID + bounding-box coordinates
and the colored rectangles/class names shown on the Roboflow screen are simply a visual representation of those numerical annotations?
If so, after I train YOLO, how can I view the actual model output/predictions and compare them with the original ground-truth annotations?
My overall goal is to compare the approaches using performance, inference speed, model size, and their ability to handle multiple solar-panel defects. Eventually, I am interested in edge/drone deployment.
I would appreciate any recommendations on how you would structure this experiment.
