How to Diagnose a Camera Out of Use Case
Computer Vision works well in the exact scenarios it was created for, and doesn't work great in other scenarios. It is important that your camera is mounted in the way the model was designed.
Bad Detections Happen
As you can see in this image, the AI thinks that this person is a vehicle. Why? Let's explore that topic.
Computer Vision Looks at the Shapes of Things
This is roughly the outline that Survail is looking for when it checks for a person. As you can see in the wrongly detected example, the distance from the subject, the angle of view, and the percentage of the person's body in view all affect that ability for the AI to tell that this is a human. To the AI, the key indicators that this is human are the presence of arms, legs, and/or a head.
Another example of this is the way that the AI tries to deal with the massive variety within the shape of things called vehicles. Vehicles can mean semis or dump trunks or coupes and nearly everything between. This means that the AI is mostly boiling the question of "Is this a vehicle?" down to two different essential questions? First, "Does it have wheels?" and secondly, since wheels are not always in view, "Is it moving, but it doesn't have arms, heads or legs?" This is basically how the AI understands a vehicle.
The vehicle model was trained by NVIDIA using 100% outdoor images, and as such it has no concept for things like rolling chairs, vacuums, or U boat shipping carts. It is going to see these things as vehicles because the fundamental questions above - "Is it moving without walking?" and "Does it have wheels?" are both true for this vacuum.
Where to find Use Case Info
Best Practices Spelled Out per Model
Each model in our Model Zoo has different requirements for what the camera must see for it to work. For example, you can detect that an object is a person equally well coming and going, but obviously if you want to do facial recognition on that person you need to (1) have the person face the camera and (2) be a lot closer to the camera to capture the details of their face. If you look at a model, you will see a list similar to the image displayed that tells you how to get the desired results with that model.
What to do if You're Outside of the Use Case
Adjust Settings
Some models are not meant for certain environments. Running our Scene Smart Zone Semantic Segmentation which labels areas as road, sidewalk, building, wall, fence, pole, traffic light, traffic sign, vegetation, terrain, and sky, obviously only works outdoors. You'll want to turn this off inside.
Move the Camera
Running facial recognition won't work on cameras that don't have faces pretty close to them. You may have to move a camera closer to get the desired result for models like facial recognition or weapon detection.
Adjust Viewport
Although it is possible for the AI to recognize people and vehicles in camera feeds that are not vertically oriented, you will have high potential that the accuracy or distance of the detection will be greatly reduced.
Camera Lens Rotation and Placement Issues
Symptom: Visual Similarity of People Detection Screenshots
In this image, we can see screenshots that are very unhelpful in determining if the video should be viewed by an operator. What's happening here is that the rotation of the camera is causing the AI to need people to be closer to the camera to determine if they are, in fact, people. By the time that the AI says with high confidence "there's a person here, take a screenshot" the person is opening the door and obscuring the view.
Vehicle Detections from the same camera do not seem to have a similar problem. as they are larger objects. However, the angle of view of these cameras would make License Plate Recognition nearly impossible. This isn't an ideal placement for LPR.
This is an much better camera placement for LPR. You want to have the camera mounted nearly perfectly behind the license plate's location with as little horizontal angle of view as possible.
Solution: Rotate Camera for Better Accuracy
In this image, we can see screenshots that after rotating the camera to have a more true-to-life angle of rotation, that the screenshots generated by Survail are more useful as it is able to detect a person earlier, thus having less interference with the door.