--- title: "TrainImageContentDetector" language: "en" type: "Symbol" summary: "TrainImageContentDetector[{img1 -> {bbox1 -> class1, ...}, ...}] trains a ContentDetectorFunction[...] based on the examples given." keywords: - object detection - object detector - image cases - image content - object recognition - object classification - computer vision - YOLO - SSD canonical_url: "https://reference.wolfram.com/language/ref/TrainImageContentDetector.html" source: "Wolfram Language Documentation" related_guides: - title: "Supervised Machine Learning" link: "https://reference.wolfram.com/language/guide/SupervisedMachineLearning.en.md" - title: "Computer Vision" link: "https://reference.wolfram.com/language/guide/ComputerVision.en.md" related_functions: - title: "ContentDetectorFunction" link: "https://reference.wolfram.com/language/ref/ContentDetectorFunction.en.md" - title: "TrainTextContentDetector" link: "https://reference.wolfram.com/language/ref/TrainTextContentDetector.en.md" - title: "ImageCases" link: "https://reference.wolfram.com/language/ref/ImageCases.en.md" - title: "Classify" link: "https://reference.wolfram.com/language/ref/Classify.en.md" - title: "NetTrain" link: "https://reference.wolfram.com/language/ref/NetTrain.en.md" - title: "FindFaces" link: "https://reference.wolfram.com/language/ref/FindFaces.en.md" - title: "FindImageShapes" link: "https://reference.wolfram.com/language/ref/FindImageShapes.en.md" - title: "ImageIdentify" link: "https://reference.wolfram.com/language/ref/ImageIdentify.en.md" --- [EXPERIMENTAL] # TrainImageContentDetector TrainImageContentDetector[{img1 -> {bbox1 -> class1, …}, …}] trains a ContentDetectorFunction[…] based on the examples given. ## Details and Options * ``TrainImageContentDetector`` attempts to train an object detector using the provided training data. * Object detection is a computer vision technique used for recognizing subimages in an image. [image] * ``TrainImageContentDetector`` returns a ``ContentDetectorFunction[…]`` that can then be applied to new input data. * The bounding boxes ``bboxi`` should be given as ``Rectangle`` objects. * The classes ``classi`` can be arbitrary expressions. * ``TrainImageContentDetector[…][image]`` can be used to detect objects in ``image`` and returns the result as a list of associations of the form ``<|"Image" -> img, "Class" -> class|>``. * ``TrainImageContentDetector`` supports the following options: | | | | | ------------------ | --------- | -------------------------------------------------------------- | | PerformanceGoal | Automatic | favor settings with specific advantages | | ProgressReporting | Automatic | how to report progress during training | | RandomSeeding | 1234 | how to seed pseudorandom generators internally | | TargetDevice | "CPU" | the target device on which to perform training | | TimeGoal | Automatic | number of seconds to train for | | ValidationSet | None | the set of data on which to evaluate the model during training | --- ## Examples (6) ### Basic Examples (1) Train a basic object detector: ```wl In[1]:= df = TrainImageContentDetector[{[image] -> {Rectangle[{89, 26}, {317, 217}] -> "heart"}, [image] -> {Rectangle[{228, 99}, {289, 158}] -> "heart"}, [image] -> {Rectangle[{86, 133}, {176, 207}] -> "heart", Rectangle[{170, 83}, {225, 130}] -> "heart"}, [image] -> {Rectangle[{147, 86}, {239, 186}] -> "heart"}}] Out[1]= ContentDetectorFunction[<|Classes -> {heart}, GridSize -> {13, 13}, Anchors -> {{0.0440562, 0.0521065}, {0.144189, 0.158656}, {0.256802, 0.421103}, {0.606371, 0.271368}, {0.751578, 0.705252}}, Net -> NetChain[<2>], DataRange -> Full, Fitting -> Fit, Version -> 1., <<2>>, Function -> NeuralFunctions`Private`file19YOLO`YOLOEvaluator, InputType -> Image, Architecture -> YOLO|>] ``` Apply the detector on a new image: ```wl In[2]:= testImage = [image]; df[testImage] Out[2]= {<|"Image" -> [image], "Class" -> "heart"|>, <|"Image" -> [image], "Class" -> "heart"|>, <|"Image" -> [image], "Class" -> "heart"|>} ``` Highlight the detection on the input image: ```wl In[3]:= HighlightImage[testImage, Legended[#BoundingBox, #Class]& /@ df[testImage, {"BoundingBox", "Class"}]] Out[3]= [image] ``` ### Options (5) #### PerformanceGoal (1) Use ``PerformanceGoal -> "Quality"`` to emphasize the quality of the result: ```wl In[1]:= training = {[image] -> {Rectangle[{48, 25}, {160, 144}] -> "apple", Rectangle[{144, 24}, {279, 144}] -> "apple"}, [image] -> {Rectangle[{84, 43}, {227, 155}] -> "strawberry"}, [image] -> {Rectangle[{60, 70}, {180, 212}] -> "strawberry"}}; TrainImageContentDetector[training, PerformanceGoal -> "Quality"] During evaluation of In[1]:= [image] ``` Use ``PerformanceGoal -> "Speed"`` to emphasize the speed of computation: ```wl In[2]:= TrainImageContentDetector[training, PerformanceGoal -> "Speed"] During evaluation of In[3]:= [image] ``` #### ProgressReporting (1) By default, progress is reported in a dynamic panel: ```wl In[1]:= training = {[image] -> {Rectangle[{48, 25}, {160, 144}] -> "apple", Rectangle[{144, 24}, {279, 144}] -> "apple"}, [image] -> {Rectangle[{84, 43}, {227, 155}] -> "strawberry"}, [image] -> {Rectangle[{60, 70}, {180, 212}] -> "strawberry"}}; In[2]:= TrainImageContentDetector[training] During evaluation of In[2]:= [image] During evaluation of In[2]:= [image] ``` Use ``ProgressReporting -> False`` to avoid displaying the progress panel: ```wl In[3]:= TrainImageContentDetector[training, ProgressReporting -> False] ``` #### TargetDevice (1) Train a detector using the default system GPU, if available: ```wl In[1]:= training = {[image] -> {Rectangle[{48, 25}, {160, 144}] -> "apple", Rectangle[{144, 24}, {279, 144}] -> "apple"}, [image] -> {Rectangle[{84, 43}, {227, 155}] -> "strawberry"}, [image] -> {Rectangle[{60, 70}, {180, 212}] -> "strawberry"}}; TrainImageContentDetector[training, TargetDevice -> "GPU"] Out[1]= ContentDetectorFunction[<|Classes -> {apple, strawberry}, GridSize -> {13, 13}, Anchors -> {{0.0440562, 0.0521065}, {0.144189, 0.158656}, {0.256802, 0.421103}, {0.606371, 0.271368}, {0.751578, 0.705252}}, Net -> NetChain[<2>], DataRange -> Full, Fitting -> Fit, <<3>>, Function -> NeuralFunctions`Private`file19YOLO`YOLOEvaluator, InputType -> Image, Architecture -> YOLO|>] ``` If a compatible GPU is not available, a message is issued: ```wl In[2]:= TrainImageContentDetector[training, TargetDevice -> "GPU"] ``` TrainImageContentDetector::trgdevmac: TargetDevice -> "GPU" is not supported on Mac OS platforms. ```wl Out[2]= [image] ``` #### TimeGoal (1) The training time can be influenced by several factors, such as the number of examples and classes: ```wl In[1]:= training = {[image] -> {Rectangle[{48, 25}, {160, 144}] -> "apple", Rectangle[{144, 24}, {279, 144}] -> "apple"}, [image] -> {Rectangle[{84, 43}, {227, 155}] -> "strawberry"}, [image] -> {Rectangle[{60, 70}, {180, 212}] -> "strawberry"}}; In[2]:= TrainImageContentDetector[training] During evaluation of In[2]:= [image] ``` Use ``TimeGoal`` to specify a target time for the training: ```wl In[3]:= TrainImageContentDetector[training, TimeGoal -> Quantity[10, "Minutes"]] During evaluation of In[4]:= [image] ``` #### ValidationSet (1) By default, only cross-validation is performed on the detector: ```wl In[1]:= training = {[image] -> {Rectangle[{48, 25}, {160, 144}] -> "apple", Rectangle[{144, 24}, {279, 144}] -> "apple"}, [image] -> {Rectangle[{84, 43}, {227, 155}] -> "strawberry"}, [image] -> {Rectangle[{60, 70}, {180, 212}] -> "strawberry"}}; In[2]:= TrainImageContentDetector[training] Out[2]= ContentDetectorFunction[<|Classes -> {apple, strawberry}, GridSize -> {13, 13}, Anchors -> {{0.0440562, 0.0521065}, {0.144189, 0.158656}, {0.256802, 0.421103}, {0.606371, 0.271368}, {0.751578, 0.705252}}, Net -> NetChain[<2>], DataRange -> Full, Fitting -> Fit, <<3>>, Function -> NeuralFunctions`Private`file19YOLO`YOLOEvaluator, InputType -> Image, Architecture -> YOLO|>] ``` Use ``ValidationSet`` to provide separate validation examples: ```wl In[3]:= validation = {[image] -> {Rectangle[{50, 74}, {139, 177}] -> "strawberry"}, [image] -> {Rectangle[{82, 44}, {132, 94}] -> "apple", Rectangle[{102, 94}, {159, 144}] -> "apple", Rectangle[{115, 8}, {170, 62}] -> "apple"}}; In[4]:= TrainImageContentDetector[training, ValidationSet -> validation] Out[4]= ContentDetectorFunction[<|Classes -> {apple, strawberry}, GridSize -> {13, 13}, Anchors -> {{0.0440562, 0.0521065}, {0.144189, 0.158656}, {0.256802, 0.421103}, {0.606371, 0.271368}, {0.751578, 0.705252}}, Net -> NetChain[<2>], DataRange -> Full, Fitting -> Fit, <<3>>, Function -> NeuralFunctions`Private`file19YOLO`YOLOEvaluator, InputType -> Image, Architecture -> YOLO|>] ``` ## See Also * [`ContentDetectorFunction`](https://reference.wolfram.com/language/ref/ContentDetectorFunction.en.md) * [`TrainTextContentDetector`](https://reference.wolfram.com/language/ref/TrainTextContentDetector.en.md) * [`ImageCases`](https://reference.wolfram.com/language/ref/ImageCases.en.md) * [`Classify`](https://reference.wolfram.com/language/ref/Classify.en.md) * [`NetTrain`](https://reference.wolfram.com/language/ref/NetTrain.en.md) * [`FindFaces`](https://reference.wolfram.com/language/ref/FindFaces.en.md) * [`FindImageShapes`](https://reference.wolfram.com/language/ref/FindImageShapes.en.md) * [`ImageIdentify`](https://reference.wolfram.com/language/ref/ImageIdentify.en.md) ## Related Guides * [Supervised Machine Learning](https://reference.wolfram.com/language/guide/SupervisedMachineLearning.en.md) * [Computer Vision](https://reference.wolfram.com/language/guide/ComputerVision.en.md) ## History * [Introduced in 2021 (13.0)](https://reference.wolfram.com/language/guide/SummaryOfNewFeaturesIn130.en.md)