Sap Se

Michael Kemelmakher

Ran Bittmann

Leonid Bobovich

Ilia Rutenburg

Laurent Boué

Yuri Arshavski

Tatiana Surazhsky

Iuliia Drozdova

Techniques for implementing blind spots into neural network models are disclosed. In some example embodiments, a computer-implemented method comprises obtaining an image captured within a field of view of an image capture device and comprising an object of a particular type occupying a particular position within the field of view, and determining a confidence value for the object based on the particular position of the object using a neural network model. The confidence value represents a likelihood that the object is an object of interest, and the neural network model is trained to generate lower confidence values for objects of the particular type when the objects of the particular type are occupying the particular position than when the objects of the particular type are not occupying the particular position.

Blind spot implementation in neural networks

In one respect, there is provided a system that may include a processor and a memory. The memory may be configured to store instructions that results in operations when executed by the processor. The operations may include: training a machine learning model by at least processing a training set with the machine learning model, the training set including at least one synthetic image that is generated by applying one or more modifications to a non-synthetic image; determining, based at least on a result of the processing of the mixed training set, that the machine learning model is unable to classify images having a specific modification; and training the machine learning model with additional training data that includes one or more additional synthetic images having the specific modification. Related methods and articles of manufacture are also disclosed.

Training machine learning models

In one respect, there is provided a system that may include a processor and a memory. The memory may be configured to store instructions that results in operations when executed by the processor. The operations may include: processing an image set with a convolutional neural network configured to detect, in the image set, a first feature and a second feature; determining a respective effectiveness of the first feature and the second feature in enabling the convolutional neural network to classify images in the image set; determining, based at least on the respective effectiveness of the first feature and the second feature, a first initial weight for the first feature and a second initial weight for the second feature; and initializing the convolutional neural network for training, the initialization of the convolutional neural network comprising configuring the convolutional neural network to apply the first initial weight and the second initial weight.

Weight initialization for machine learning models

In one respect, there is provided a system that may include a processor and a memory. The memory may be configured to store instructions that results in operations when executed by the processor. The operations may include: generating a concatenated feature map set by at least combining a first feature map set and a second feature map set, the first feature map set and the second feature map set each indicating one or more occurrences of a feature within an image, the first feature map set having a different scale than the second feature map set; and classifying, by a convolutional neural network, an image, the classifying of the image being based at least on the concatenated feature map set. Related methods and articles of manufacture are also disclosed.

Small object detection

Methods, systems, and computer-readable storage media for receiving a set of frames, each frame being provided as a digital image that depicts a portion of an event and a logo associated with a brand, for each frame in the set of frames, and for each pixel in a frame: determining a weight of the pixel based on a distribution assigned to the frame, providing a quality of the logo depicted in the frame, and calculating a pixel quotient based on the weight and the quality, for each frame in the set of frames: determining a frame quotient at least partially based on a sum of all pixel quotients for the frame, and determining an impact indicator for the logo based on a total size of digital media comprising the set of frames and a sum of frame quotients of the frames in the set of frames.

Growing community of inventors

Hod Hasharon, Israel

Michael Kemelmakher