EFFICIENT LOSSY CONTOUR CODING USING SPATIO-TEMPORAL CONSISTENCY Chaumont Marc, Pateux St´ ephane and Nicolas Henri IRISA, Campus de Beaulieu, 35042 Rennes, France Email: Marc.Chaumont@irisa.fr ABSTRACT This paper deals with the problem of lossy video object shape coding. This problem belongs to the area of video object-based coder. Our approach proposes to exploit spatio temporal correla- tion. This is done by considering the contour evolution and dealing with occultation phenomenon. The novelty lies in a fine mapping of consecutive contours and in the contour padding in occulted parts. Once a group of contours is processed, wavelet transforma- tions with a predictive IPB (Intra, simple Prediction, Bidirectional prediction) scheme is used to encode it. Experimental results show significant improvement at very low bit rate. Moreover, the bit- stream is fully progressive. 1. INTRODUCTION In object-based video coder such as in MPEG4 [1], shape information has to be coded. Efficient coding techniques have been studied to decrease the bit-rate overhead of shape [2][3]. We generally classify those shape coding techniques in two categories, the bitmap-based coding or the contour- based coding. The bitmap-based coding approaches (MMR: Modified- Modified Read, CAE: Context-Based Arithmetic Encoding, quad-tree...) rely on a bileveling coding technique. As an example, CAE technique [4], uses an arithmetic technique to code each pixel knowing their surrounding context. Some of these approaches take into account temporal redundancy, for instance MPEG4 has developed a solution based on hi- erarchical BAB (Binary Alpha Blocks), CAE technique and temporal prediction. MPEG4 CAE is the present reference solution for lossy shape coding. The problem of this solu- tion is bloc artifacts at low bit-rate, which is inherent to the adopted modeling. On the other side, contour-based coding approaches (Free- man, polygonal, B-spline...) only describe the contour. Free- man technique codes a chain of directional displacements. Polygonal and B-spline approach define contours by means of lines segments and respectively control points. Yoshida and al. [5] have proposed a lossy contour based solution equaling MPEG4 CAE’s performance at low bit rate. They took into account the temporal information by simultaneously coding a group of consecutive contours be- longing to the same video object. Our approach is inspired by their solution. The main novelty is that we are improving temporal stability by working only with real shape contour. This is done by taking into account contour’s part which really belong to the object (see Fig.1). Further, a better tem- poral mapping is proposed exploiting motion compensation. Thus, we could deals with non rigid object. (a) Initial segmentation map (b) Partial outer contour Fig. 1. Extraction of the apparent contour of an object 2. EXTRACTION AND ALIGNMENT OF A CONTOUR BELONGING TO A GROUP OF FRAME Our objective is to obtain for all contour’s points, the corre- sponding position all along time in the purpose of extract- ing the evolution surface of the contour. This surface could be represented by two spatio-temporal planes giving a point position knowing indices on the contour and know- ing the frame number (see Fig. 8). Thus, a bijective mapping of consecutive contours along time should be found. The major problem is that the contour of a video object do not have a constant length throughout time. This is due to deformations, occultation with other ob- jects or image borders. As consequence the contour points may be mapped to zero, one or several points. To solve this problem, points must be artificially added in such a way that all points can be mapped one to one. The different steps to reach our objective are first the contour extraction and the definition of a list of positions for each contour belonging to the group of frame, second, an adapted mapping between consecutive contours, third, the contours alignment thanks to the computation of an uni- versal abscissa and finally, a spatio-temporal padding to fill missing positions.