AWS Sagemaker Multiple Object Detection in Image Recognition / Classification

Question

Does anyone know if any SageMaker built-in algorithm supports multiple object detection in image recognition? I am thinking something like multi-label image training and detection / inference.

Thus, can we:

a) train using multi-label images

and/or

b) infer multiple objects from images (sort of like AWS Rekognition but with custom labels and training / transfer learning).

Also, I know that the doc for SageMaker Image Classification Algorithm says "takes an image as input and classifies it into one of multiple output categories".

Any recommendations are also welcome.

Answer 1

There is a new built-in algorithm released with Amazon Sagemaker today for object detection. Based on the documentation, Amazon SageMaker Object Detection uses the Single Shot multibox Detector (SSD) algorithm. The response from the inference contains an array consists of a predicted class label of the object detected, associated confidence score and bounding box co-ordinates.

Answer 2

UPDATE: Checkout Object Detection Algorithm in SageMaker

---

SageMaker's image classification algorithm takes in an image and returns a collection of class and probability.
To classify multiple objects in an image, you need to pre-process the image to detect objects, resulting in multiple smaller images with one object per image, which you'll provide as input to the image classification algorithm.

Typically object detection is an easier task than object identification/classification.