After training a computer vision model, you may ask yourself “how does this model version compare to the last version?” Metrics like mAP and accuracy give you an aggregate view across all images, but you may want to be more precise and probe how your model does on particular images.
Direct, visual comparison of results can help you better determine how your previous and current models perform, and help you decide next steps on how to improve your model.
In this walkthrough we will use the Model Comparison Visualization block in Roboflow Workflows to compare predictions from two object detection models.
Here is an example of the block running on an image using models that aim to identify individual wood ends for use in a counting application:
The red marks areas that our first model did not detect but the second one did. This tells us that our second model is significantly more effective at identifying wood ends.
By the end of this tutorial, you will be able to visualize where each model detects objects differently, helping you make a more informed decision about which model best meets your needs. Let’s get started!
Why Use Model Comparison Visualization?
The Model Comparison Visualization block in Roboflow Workflows allows you to see exactly which areas in an image one model finds that the other model does not. Here’s where it can be helpful:
- Two Models, Similar Benchmarks: Even if two models have similar performance (for example, an mAP of 0.85 vs. 0.84), the differences in what each model detects may be important. One model may catch edge cases better, or could generalize better to certain scenarios.
- Compare an Older Model to a Newer Iteration: After retraining a model with more data or improved labeling, it’s important to see how predictions differ on real-world images. You might discover the new model identifies additional objects or eliminates false positives.
- Different Data Requirements: You could have one model trained on less data than another. While both might show good metrics, seeing how they perform visually on the same image can help you decide which one is better for your use case.
- Faster vs. More Accurate: If you have a faster but potentially less accurate model vs. a slower, more accurate one, you can use visual comparison to see how each trade-off manifests on real images.
Step #1: Create a Workflow
First, navigate to the “Workflows” tab in the left sidebar and click “Create Workflow”.
Step #2: Add Your Model Blocks
Click on “Add Block" and choose whichever model you’d like to compare. In my case, I’m using an object detection model from Roboflow Universe. We want to run inference using two different models on the same input image.
To do this, branch from the first model block by hovering over the Model block for your first model and click the branch icon (a small node connector).
This lets you add a second block in parallel to the first model. Below you’ll see my two models in parallel.
Step #3: Add the Model Comparison Visualization Block
With both models added, we can now introduce the Model Comparison Visualization block to visualize the difference in predictions.
Click on the plus (+) icon under either of the model blocks to create a new block. Select Model Comparison Visualization from the list of available blocks.
This block compares the predictions made by Model A and Model B. Let’s set it up:
- Predictions A: Set this to model_a -> predictions.
- Predictions B: Set this to model_b -> predictions.
Step #4: Test Your Workflow
With your Workflow set up, it’s time to test. Click “Save” and then click on “Test Workflow”.
Upload your image to be fed to both models and then compared onto a single image in the output.
- Green areas where Model A “wood-ends/1” (older model) made predictions that Model B missed.
- Red areas where Model B “wood-ends/8” (newer model) predicted ends that Model A did not detect.
- Black where neither model made a prediction.
As you can see, the model comparison block showed us that the newer version of our model is significantly more accurate in identifying wooden ends.
While we used green and red above, you can use custom colors. You can change the colours in the additional properties section of the model comparison block:
Conclusion
The Model Comparison Visualization block in Roboflow Workflows is a powerful way to visually inspect the differences between two sets of predictions.
This is useful for checking whether a new model truly outperforms an older iteration in real-world applications, comparing models trained with different data or settings, and quickly identifying missed detections or false positives in a color-coded way.
With these skills, you can confidently assess how each model does in real scenarios and make data-driven decisions about which model to deploy in production.
Check out the Roboflow Workflow Gallery to see documentation and also our Workflows Blocks Gallery to explore more blocks and help you make a wide range of computer vision pipelines.
Happy building!
Cite this Post
Use the following entry to cite this post in your research:
Asfandiyar Khan. (Jan 24, 2025). How to Visually Compare Computer Vision Models. Roboflow Blog: https://blog.roboflow.com/compare-computer-vision-models-visually/
Discuss this Post
If you have any questions about this blog post, start a discussion on the Roboflow Forum.