Partial versus Early Packet Video Discard Ahmed Mehaoua University of Cambridge 10 Downing Street Cambridge, United Kingdom am296@ccsr.cam.ac.uk Raouf Boutaba University of Toronto 10 King’s College Road Toronto, Canada rboutaba@nal.utoronto.ca Youssef Iraqi University of Montreal C.P. 6128, succ. Centre-Ville Montreal, Canada iraqi@iro.umontreal.ca ABSTRACT In this paper, we propose and compare two video slice- based discard schemes, namely adaptive-PSD and Adaptive-ESD, for the transmission of MPEG video streams overATM best effort services. The schemes perform adaptive and selective cell burst discard at the level of MPEG video slices and intelligently adjusts drop policies to switch buffer occupancy and video cell payload types. In comparison to previous approaches, the performance evaluation have shown a significant reduction of the bad throughput crossing the network and a better protection of critical Intra- and Predictive-coded pictures at both cell and video slice levels. Keywords: ATM, Best Effort, Packet Video, MPEG, Cell Discard. I. INTRODUCTION With increasing interest in the transmission of MPEG- compressed video streams over unreliable ATM best effort services (ABR, UBR+), efficient video-oriented packet dropping mechanisms have to be designed which attempt to gracefully control picture quality degradation during network congestion. These video applications will extensively make use of MPEG video compression standards to save network resources. MPEG defines a video stream as a hierarchy of data structures ordered by increasing spatial size : pixel, 8x8 pixel Block, 16x16 pixel MacroBlock, Slice, Frame, Group of Pictures and Sequence [1]. Two of them have significant impacts on the decoding/displaying process and thereby on the picture quality perceived by the end users. Video slice is the main coding processing unit in MPEG. Coding and decoding of blocks and macroblocks are feasible only when all the pixels of a slice are available. Besides, encoding of a slice is done independently from its adjacent slices, making it the smallest autonomous decoding unit. Consequently, it serves as resynchronization point in case of problems. Frame or picture is the basic unit of display. Three picture types may be present in a MPEG video stream. They differ from the encoding method used: Intra- coded (I) picture, Predictive-coded (P) picture and Bidirectionally predictive-coded (B) picture. I- and P- encoded pictures are essential and have to be preserved from corruption during transmission. Due to error propagation at the decoding layer, a corrupted or non- available reference picture (e.g. I- or P-frame) leads on perceptible picture degradation. I-frame impairments will affect all the subsequent frames on the same Group Of Picture (GOP). Similarly, the impairment of P-frames will affect the following P- and B-frames until the next I-frame. Only B-frame impairments have no adverse effects on other frames. According to the above statements, three obvious remarks stand out. First, the smallest transmission data unit is rather video slice than ATM cell, AAL PDU or MPEG multiplex packet (e.g. Transport Stream or Program Stream). Secondly, in situation of congestion, dropping video cells indiscriminately can cause serious degradation in picture quality. Intra- and Predictive frames have to be better protected from errors during transmission. Finally, without intelligent FEC and Error concealment mechanisms at destination, forwarding partially corrupted video slices is wasteful and may even worsen the congestion in the upstream nodes. Thus, the question arises which is the best cell dropping policy that ensure the highest network bandwidth utilization while minimizing the video slice loss probability at the application layer. To address this problem, we present and compare the performance of two packet video drop policies for use with ATM best effort services (i.e. ABR, UBR+, GFR). The schemes are referred to Adaptive and Partial video Slice Discard (A-PSD) and Adaptive and Early video Slice Discard (A-ESD). The paper is organized as follows. In section 2, we briefly review some previous works in this area.. Section 3 is devoted to the description of the two proposed video cell discarding schemes. Section 4 introduces the network simulation model and discuss the performance results. Finally, we conclude in Section 5.