Model-based enhancement of lighting conditions in image sequences

Illumination variability has a considerable influence on the performance of computer vision algorithms or video coding methods. The efficiency and robustness of these algorithms can be significantly improved by removing the undesired effects of changing illumination. In this paper, we introduce a 3-D model-based technique for estimating and manipulating the lighting in an image sequence. The current scene lighting is estimated for each frame exploiting 3-D model information and by synthetic re-lighting of the original video frames. To provide the estimator with surface normal information, the objects in the scene are represented by 3-D shape models and their motion and deformation are tracked over time using a model-based estimation method. Given the normal information, the current lighting is estimated with a linear algorithm of low computational complexity using an orthogonal set of light maps. This results in a small set of parameters which efficiently represent the scene lighting. In our experiments, we demonstrate how this representation can be used to create video sequences with arbitrary new illumination. When encoding a video sequence for transmission over a network, significant coding gains can be achieved when removing the time varying lighting effects prior to encoding. Improvements of up to 3 dB in PSNR are observed for an illumination-compensated sequence.