The trained object detection model is not working on the new test images. It works only for the training dataset. Anybody else facing the same issue?
Hi,
You might need more data to train your model. The issue of your model only being able to detect training images might suggest that your model needs more data in order to recognize properly.
Some other things to consider:
- Roboflow’s automatic dataset splits usually work fine, but I’ve found that it sometimes sorts an unrepresentative sample of my images into the train/valid/test dataset. Have you tried checking how your images are being split up?
- I’ve also found that detections on new images usually have a lower confidence level. Try checking your minimum confidence settings when inferring on your test dataset.
1 Like
Great response @leo!
As suggested, @Navaneetha_Krishnan, you’ll likely need more images in your dataset that are representative of the images/video you are testing your model on for it to have better performance.
A lower confidence level setting for inference will also be helpful, as well as checking the Class Balance for your train and valid sets to ensure that each class is represented in both splits:
- You can use the Dataset Search feature to Filter by Class: Launch: Dataset Search
I recommend trying the Active Learning approach, too: What is Active Learning? | Active Learning - Roboflow