Boosting object detection performance in videos

A computer program product for training a special object detector to distinguish a foreground object appearing in a sequence of frames for a target domain, the sequence of frames depicting motion of the foreground object in a non-uniform background of the sequence, each frame in the sequence before a last frame in the sequence corresponding to a time earlier than a time to which an immediately next frame in the sequence corresponds, the computer program product comprising:

• a computer readable hardware storage device that is not a signal or a medium/wave;
• first program instructions to detect the foreground object in a high-confidence subwindow of an initial frame of the sequence, wherein to detect the foreground object comprises to compute a measure of confidence that the high-confidence subwindow comprises the foreground object and to determine that the measure of confidence exceeds a specified confidence threshold;
• second program instructions to track the foreground object in respective positive subwindows of a plurality of subsequent frames appearing after the initial frame in the sequence of frames, wherein the subsequent frames are within a specified short period of time;
• and
• third program instructions to feed the positive subwindows to a learning engine for use in training the special object detector to detect the foreground object in the target domain, wherein the positive subwindows include the subwindow of the initial frame and the respective subwindows of the plurality of subsequent frames,
• wherein the computer program product further comprises:
• fourth program instructions to randomly select a negative subwindow in each window of a set of windows pertaining to the target domain, wherein the negative subwindow in each window of the set of windows has an aspect ratio differing from the aspect ratio of the high-confidence subwindow, and wherein the negative subwindow in each window of the set of windows does not include the foreground object;
• and
• fifth program instructions to resize the negative subwindow in each window of the set of windows to become congruent to the high-confidence subwindow, and
• wherein the first program instruction, the second program instructions, the third program instructions, the fourth program instructions, and the fifth program instructions are stored on the computer readable hardware storage device for execution by one or more processors of a computer system.

A computer system for training a special object detector to distinguish a foreground object appearing in a sequence of frames for a target domain, the sequence of frames depicting motion of the foreground object in a non-uniform background, each frame in the sequence before a last frame in the sequence corresponding to a time earlier than a time to which an immediately next frame in the sequence corresponds, the computer system comprising:

• one or more processors;
• a computer readable memory;
• a computer readable storage device;
• first program instructions to detect the foreground object in a high-confidence subwindow of an initial frame of the sequence, wherein to detect the foreground object comprises to compute a measure of confidence that the high-confidence subwindow comprises the foreground object and to determine that the measure of confidence exceeds a specified confidence threshold;
• second program instructions to track the foreground object in respective positive subwindows of a plurality of subsequent frames appearing after the initial frame in the sequence of frames, wherein the subsequent frames are within a specified short period of time;
• and
• third program instructions to feed the positive subwindows to a learning engine for use in training the special object detector to detect the foreground object in the target domain, wherein the positive subwindows include the subwindow of the initial frame and the respective subwindows of the plurality of subsequent frames,
• wherein the computer program system further comprises:
• fourth program instructions to randomly select a negative subwindow in each window of a set of windows pertaining to the target domain, wherein the negative subwindow in each window of the set of windows has an aspect ratio differing from the aspect ratio of the high-confidence subwindow, and wherein the negative subwindow in each window of the set of windows does not include the foreground object;
• and
• fifth program instructions to resize the negative subwindow in each window of the set of windows to become congruent to the high-confidence subwindow, and
• wherein the first program instruction, the second program instructions, the third program instructions, the fourth program instructions, and the fifth program instructions are stored on the computer readable computer readable storage device for execution by one or more processors via the computer readable memory.