diff --git a/docs/docs/plus/index.md b/docs/docs/plus/index.md index 7d92fb531..687d01cce 100644 --- a/docs/docs/plus/index.md +++ b/docs/docs/plus/index.md @@ -9,6 +9,34 @@ With a subscription, and at each annual renewal, you will receive 12 model train Information on how to integrate Frigate+ with Frigate can be found in the [integrations docs](/integrations/plus). +## Improving your model + +You may find that Frigate+ models result in more false positives initially, but by submitting true and false positives, the model will improve. Because a limited number of users submitted images to Frigate+ prior to this launch, you may need to submit several hundred images per camera to see good results. With all the new images now being submitted, future base models will improve as more and more users (including you) submit examples to Frigate+. + +False positives can be reduced by submitting **both** true positives and false positives. This will help the model differentiate between what is and isn't correct. You should aim for a target of 80% true positive submissions and 20% false positives across all of your images. If you are experiencing false positives in a specific area, submitting true positives for any object type near that area in similar lighting conditions will help teach the model what that area looks like when no objects are present. + +You may find that it's helpful to lower your thresholds a little in order to generate more false/true positives near the threshold value. For example, if you have some false positives that are scoring at 68% and some true positives scoring at 72%, you can try lowering your threshold to 65% and submitting both true and false positives within that range. This will help the model learn and widen the gap between true and false positive scores. + +Note that only verified images will be used when training your model. Submitting an image from Frigate as a true or false positive will not verify the image. You still must verify the image in Frigate+ in order for it to be used in training. + +In order to request your first model, you will need to have annotated and verified at least 10 images. Each subsequent model request will require that 10 additional images are verified. However, this is the bare minimum. For the best results, you should provide at least 100 verified images per camera. Keep in mind that varying conditions should be included. You will want images from cloudy days, sunny days, dawn, dusk, and night. + +As circumstances change, you may need to submit new examples to address new types of false positives. For example, the change from summer days to snowy winter days or other changes such as a new grill or patio furniture may require additional examples and training. + +## Properly labeling images + +For the best results, follow the following guidelines. + +**Label every object in the image**: It is important that you label all objects in each image before verifying. If you don't label a car for example, the model will be taught that part of the image is _not_ a car and it will start to get confused. + +**Make tight bounding boxes**: Tighter bounding boxes improve the recognition and ensure that accurate bounding boxes are predicted at runtime. + +**Label the full object even when occluded**: If you have a person standing behind a car, label the full person even though a portion of their body may be hidden behind the car. This helps predict accurate bounding boxes and improves zone accuracy and filters at runtime. + +**`amazon`, `ups`, and `fedex` should label the logo**: For a Fedex truck, label the truck as a `car` and make a different bounding box just for the Fedex logo. If there are multiple logos, label each of them. + +![Fedex Logo](/img/plus/fedex-logo.jpg) + ## Frequently asked questions ### Are my models trained just on my image uploads? How are they built? @@ -17,7 +45,7 @@ Frigate+ models are built by fine tuning a base model with the images you have a ### What is a training credit and how do I use them? -Essentially, `1 training credit = 1 trained model`. When you have uploaded, annotated, and verified additional images and you are ready to train your model, you will submit a model request which will use one credit. The model that is trained will utilize all of the verified images in your account. +Essentially, `1 training credit = 1 trained model`. When you have uploaded, annotated, and verified additional images and you are ready to train your model, you will submit a model request which will use one credit. The model that is trained will utilize all of the verified images in your account. When new base models are available, it will require the use of a training credit to generate a new user model on the new base model. ### Are my video feeds sent to the cloud for analysis when using Frigate+ models? @@ -109,31 +137,3 @@ When using Frigate+ models, Frigate will choose the snapshot of a person object `amazon`, `ups`, and `fedex` labels are used to automatically assign a sub label to car objects. ![Fedex Attribute](/img/plus/attribute-example-fedex.jpg) - -## Properly labeling images - -For the best results, follow the following guidelines. - -**Label every object in the image**: It is important that you label all objects in each image before verifying. If you don't label a car for example, the model will be taught that part of the image is _not_ a car and it will start to get confused. - -**Make tight bounding boxes**: Tighter bounding boxes improve the recognition and ensure that accurate bounding boxes are predicted at runtime. - -**Label the full object even when occluded**: If you have a person standing behind a car, label the full person even though a portion of their body may be hidden behind the car. This helps predict accurate bounding boxes and improves zone accuracy and filters at runtime. - -**`amazon`, `ups`, and `fedex` should label the logo**: For a Fedex truck, label the truck as a `car` and make a different bounding box just for the Fedex logo. If there are multiple logos, label each of them. - -![Fedex Logo](/img/plus/fedex-logo.jpg) - -## Improving your model - -You may find that Frigate+ models result in more false positives initially, but by submitting true and false positives, the model will improve. This may be because your cameras don't look quite enough like the user submissions that were used to train the base model. Over time, this will improve as more and more users (including you) submit examples to Frigate+. - -False positives can be reduced by submitting **both** true positives and false positives. This will help the model differentiate between what is and isn't correct. - -You may find that it's helpful to lower your thresholds a little in order to generate more false/true positives near the threshold value. For example, if you have some false positives that are scoring at 68% and some true positives scoring at 72%, you can try lowering your threshold to 65% and submitting both true and false positives within that range. This will help the model learn and widen the gap between true and false positive scores. - -Note that only verified images will be used when training your model. Submitting an image from Frigate as a true or false positive will not verify the image. You still must verify the image in Frigate+ in order for it to be used in training. - -In order to request your first model, you will need to have annotated and verified at least 10 images. Each subsequent model request will require that 10 additional images are verified. However, this is the bare minimum. For the best results, you should provide at least 100 verified images per camera. Keep in mind that varying conditions should be included. You will want images from cloudy days, sunny days, dawn, dusk, and night. - -As circumstances change, you may need to submit new examples to address new types of false positives. For example, the change from summer days to snowy winter days or other changes such as a new grill or patio furniture may require additional examples and training.