A caption detection system wherein all detected caption boxes over time for one caption area are identical, thereby reducing temporal instability and inconsistency. This is achieved by grouping candidate pixels in the 3D spatiotemporal space and generating a 3D bounding box for one caption area. 2D bounding boxes are obtained by slicing the 3D bounding boxes, thereby reducing temporal instability as all 2D bounding boxes corresponding to a caption area are sliced from one 3D bounding box and are therefore identical over time.