![Evaluation
Evaluation
Test Methodology
Test videos: JVET test sequences5 and professionally produced UHD HDR cinematic con-
tent6 having typical multi-scene content
System: Dual-processor server with Intel Xeon Gold 5218R (80 cores, 2.10 GHz)
Benchmark algorithm: default shot detection algorithm in x265
T1 = 50 and T2 = 10 for the proposed algorithm; determined experimentally
Metrics: accuracy, precision, recall,7 and F-measure8
5
Jill Boyce et al. JVET-J1010: JVET common test conditions and software reference configurations. July 2018.
6
M. H. Pinson. “The Consumer Digital Video Library [Best of the Web]”. In: IEEE Signal Processing Magazine 30.4 (2013), pp. 172–174. doi:
10.1109/MSP.2013.2258265.
7
Markus Junker, Rainer Hoch, and Andreas Dengel. “On the Evaluation of Document Analysis Components by Recall, Precision, and Accuracy”. In: (Apr.
2000). doi: 10.1109/ICDAR.1999.791887.
8
Sasaki Yutaka. “The truth of the F-measure”. In: https://www.toyota-ti.ac.jp/Lab/Denshi/COIN/people/yutaka.sasaki/F-measure-YS-26Oct07.pdf. 2007.
Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 17](https://image.slidesharecdn.com/icipshotdetectionpresentation2-220220181028/85/IEEE-ICIP-22-Efficient-Content-Adaptive-Feature-based-Shot-Detection-for-HTTP-Adaptive-Streaming-17-320.jpg)
IEEE ICIP'22:Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming
- 1. Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming Vignesh V Menon, Hadi Amirpour, Mohammad Ghanbari, Christian Timmerer Christian Doppler Laboratory ATHENA, Institute of Information Technology (ITEC), University of Klagenfurt, Austria 19-22 September 2021 Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 1
- 2. Outline 1 Introduction 2 Shot detection 3 Proposed Algorithm 4 Evaluation 5 Conclusions and Future Directions Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 2
- 3. Introduction Introduction Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 3
- 4. Introduction Introduction Background of HTTP Adaptive Streaming (HAS)1 Source: https://bitmovin.com/adaptive-streaming/ Why Adaptive Streaming? Adapt for a wide range of devices Adapt for a broad set of Internet speeds What HAS does? Each source video is split into segments Encoded at multiple bitrates, resolutions, and codecs Delivered to the client based on the device capability, network speed etc. 1 A. Bentaleb et al. “A Survey on Bitrate Adaptation Schemes for Streaming Media Over HTTP”. In: IEEE Communications Surveys Tutorials 21.1 (2019), pp. 562–585. doi: 10.1109/COMST.2018.2862938. Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 4
- 5. Introduction Introduction Multi-shot encoding framework for VoD HAS applications2 Input Video Shot Detection Shot Encodings Video Quality Measure Convex Hull Determination Encoding Set Generation Multi-shot Encoding Encoded Shots Bitrate Quality Pairs Bitrate Resolution Pairs Target Encoding Set 2 Venkata Phani Kumar M, Christian Timmerer, and Hermann Hellwagner. “MiPSO: Multi-Period Per-Scene Optimization For HTTP Adaptive Streaming”. In: 2020 IEEE International Conference on Multimedia and Expo (ICME). 2020, pp. 1–6. doi: 10.1109/ICME46284.2020.9102775. Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 5
- 6. Shot detection Shot detection Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 6
- 7. Shot detection Shot Detection The boundaries between video shots are commonly known as shot transitions or shot-cuts. The act of segmenting a video sequence into shots is called shot detection. Objective: Detect the first picture of each shot and encode it as an Instantaneous Decoder Refresh (IDR) frame. Encode the subsequent frames of the new shot based on the first one via motion compen- sation and prediction.3 3 J.-R Ding and Jar-Ferr Yang. “Adaptive group-of-pictures and scene change detection methods based on existing H.264 advanced video coding information”. In: Image Processing, IET 2 (May 2008), pp. 85 –94. doi: 10.1049/iet-ipr:20070014. Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 7
- 8. Shot detection Shot Detection Shot transitions can be present in two ways: hard shot-cuts gradual shot transitions The detection of gradual changes is much more difficult owing to the fact it is difficult to determine the change in the visual information in a quantitative format. Note 1 Ratio of IDR frames to non-IDR frames is skewed, i.e, uneven distribution. 2 Missed shot-cut detections and wrong IDR placements cause low compression efficiency, i.e., cost of error is large. Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 8
- 9. Proposed Algorithm Proposed Algorithm Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 9
- 10. Proposed Algorithm Phase 1: Feature Extraction Proposed Algorithm Phase 1: Feature Extraction Compute texture energy per Coding Tree Unit (CTU) A DCT-based energy function is used to determine the block-wise feature of each frame defined as: Hk = w X i=1 h X j=1 e|( ij wh )2−1| |DCT(i − 1, j − 1)| (1) where w and h are the width and height of the block, and DCT(i, j) is the (i, j)th DCT component when i + j > 2, and 0 otherwise. The energy values of CTUs in a frame is averaged to determine the energy per frame.4 4 Michael King, Zinovi Tauber, and Ze-Nian Li. “A New Energy Function for Segmentation and Compression”. In: July 2007, pp. 1647–1650. doi: 10.1109/ICME.2007.4284983. Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 10
- 11. Proposed Algorithm Phase 1: Feature Extraction Proposed Algorithm Phase 1: Feature Extraction Figure: Hk of Tears of Steel sequence. Black circles denote the regions of shot transitions. Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 11
- 12. Proposed Algorithm Phase 1: Feature Extraction Proposed Algorithm Phase 1: Feature Extraction hk: Mean Squared Error (MSE) of the CTU level energy values of frame k to that of the previous frame k − 1, normalized to Hk. hk = PM i=1(Hk(i) − Hk−1(i))2 MHk (2) where M denotes the number of CTUs in frame k. : gradient of h per frame, given by: k = hk−1 − hk hk−1 (3) Note If hk = 0, kth frame is a duplicate of (k − 1)th frame. Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 12
- 13. Proposed Algorithm Phase 2: Successive Elimination Algorithm Proposed Algorithm Phase 2: Successive Elimination Algorithm Step 1: while Parsing all video frames do if k T1 then k ← IDR-frame, a new shot. else if k ≤ T2 then k ← P-frame or B-frame, not a new shot. T1 , T2 : maximum and minimum threshold for k Note The frames are classified into three categories in this step: 1 a new shot 2 not a new shot 3 not decided In the next steps of the algorithm, only frames of category (3) are considered. Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 13
- 14. Proposed Algorithm Phase 2: Successive Elimination Algorithm Proposed Algorithm Phase 2: Successive Elimination Algorithm f : video fps Q : set of frames where T1 ≥ T2 q0: current frame number in the set Q q−1: previous frame number in the set Q q1: next frame number in the set Q Step 2: while Parsing Q do if q0 − q−1 f and q1 − q0 f then q0 ← IDR-frame, a new shot. Eliminate q0 from Q. Step 3: while Parsing Q do if q0 − q−1 f and q1 − q0 ≤ f then compare q0 with q when q is from the subset of Q where q1 − q0 ≤ f Frame q with the highest value ← IDR-frame, a new shot. Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 14
- 15. Proposed Algorithm Phase 2: Successive Elimination Algorithm Proposed Algorithm Working Example Table: Step 1. Frame Hk 33 52162 21.68 54 52119 13.51 65 52625 19.21 86 52038 10.12 97 52499 17.34 161 47790 11.53 833 48644 11.49 1409 40367 14.51 1665 35321 19.93 1686 40463 10.72 1889 38475 12.16 2205 37218 10.08 2536 35793 10.49 Table: Step 2. Frame Hk q0 − q−1 q1 − q0 33 52162 21.68 33 21 54 52119 13.51 21 11 65 52625 19.21 11 21 86 52038 10.12 21 11 97 52499 17.34 11 64 161 47790 11.53 64 672 833 48644 11.49 672 576 1409 40367 14.51 576 256 1665 35321 19.93 256 21 1686 40463 10.72 21 203 1889 38475 12.16 203 316 2205 37218 10.08 316 331 2536 35793 10.49 331 - Table: Step 3. Frame Hk q0 − q−1 q1 − q0 33 52162 21.68 33 21 54 52119 13.51 21 11 65 52625 19.21 11 21 86 52038 10.12 21 11 97 52499 17.34 11 64 1665 35321 19.93 256 21 1686 40463 10.72 21 203 2536 35793 10.49 331 - This example uses FunOnTheRiver (24 fps) test sequence. Detected frames to be encoded as IDR-frames in each step are: Step 1: - Step 2: 161, 833, 1409, 1889, 2205 Step 3: 33, 1665, 2536 Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 15
- 16. Evaluation Evaluation Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 16
- 17. Evaluation Evaluation Test Methodology Test videos: JVET test sequences5 and professionally produced UHD HDR cinematic con- tent6 having typical multi-scene content System: Dual-processor server with Intel Xeon Gold 5218R (80 cores, 2.10 GHz) Benchmark algorithm: default shot detection algorithm in x265 T1 = 50 and T2 = 10 for the proposed algorithm; determined experimentally Metrics: accuracy, precision, recall,7 and F-measure8 5 Jill Boyce et al. JVET-J1010: JVET common test conditions and software reference configurations. July 2018. 6 M. H. Pinson. “The Consumer Digital Video Library [Best of the Web]”. In: IEEE Signal Processing Magazine 30.4 (2013), pp. 172–174. doi: 10.1109/MSP.2013.2258265. 7 Markus Junker, Rainer Hoch, and Andreas Dengel. “On the Evaluation of Document Analysis Components by Recall, Precision, and Accuracy”. In: (Apr. 2000). doi: 10.1109/ICDAR.1999.791887. 8 Sasaki Yutaka. “The truth of the F-measure”. In: https://www.toyota-ti.ac.jp/Lab/Denshi/COIN/people/yutaka.sasaki/F-measure-YS-26Oct07.pdf. 2007. Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 17
- 18. Evaluation Evaluation Experimental Results Table: Shot detection results Video Actual Benchmark algorithm Proposed algorithm shot-cuts Accuracy Precision Recall F-measure Accuracy Precision Recall F-measure BigBuckBunny 10 99.88% 100.00% 80.00% 88.89% 100.00% 100.00% 100.00% 100.00% Dinner 4 99.89% 100.00% 75.00% 85.71% 99.89% 100.00% 75.00% 85.71% FoodMarket4 2 99.72% - 0% - 99.86% 100.00% 50.00% 66.67% sintel trailer 14 99.86% 100.00% 85.71% 92.31% 99.93% 100.00% 92.86% 96.30% snow mnt 3 99.47% - 0% - 99.65% 100.00% 33.33% 50.00% Tears of Steel 13 99.93% 100.00% 92.31% 96.00 % 100.00% 100.00% 100.00% 100.00% Busy City 11 99.64% 50.00% 18.18% 26.67% 99.87% 100.00% 63.64% 77.78% FunOnTheRiver 12 99.60% 0% 0% - 99.80% 85.71% 50.00% 63.16% Remarks 1 Actual shot-cuts: the ground truth, i.e., the number of real shot transitions in the considered test videos determined manually. 2 Recall rate of the proposed algorithm is 25% better than the benchmark algorithm. 3 F-measure of the proposed algorithm is 20% higher compared to the benchmark algorithm. Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 18
- 19. Evaluation Evaluation Experimental Results Table: Detection rate statistics of the algorithms Algorithm TPR FPR Benchmark 53.62% 0.03% Proposed 78.26% 0.01% Runtime per frame: 0.1% of the total time taken for encoding each frame. The algorithm needs to be run only once for a video. The decisions made can be used for all remaining representations in HAS applications. Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 19
- 20. Conclusions and Future Directions Conclusions and Future Directions Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 20
- 21. Conclusions and Future Directions Conclusions Proposed a shot detection algorithm as a feature-based pre-processing step for x265-based HEVC encoding in VoD HAS applications. Identified a DCT-based energy function as a feature to determine shot cuts. Proposed a successive elimination algorithm to remove the false detections during gradual shot transitions. The proposed algorithm gives better-balanced shot detections compared to the benchmark algorithm. Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 21
- 22. Conclusions and Future Directions Future Directions We can extend the work in this paper to compute the relative complexity of the shots to that of the entire video sequence using the feature metric and predict the ideal bitrate per resolution for each shot. As an extension of this work, more encoding parameter decisions like optimal block parti- tioning, quantization offsets can be predicted. This work can be extended to support more recent codecs e.g., VVC. Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 22
- 23. Conclusions and Future Directions Q A Thank you for your attention! Vignesh V Menon (vignesh.menon@aau.at) Hadi Amirpour (hadi.amirpourazarian@aau.at) Mohammad Ghanbari (ghan@essex.ac.uk) Christian Timmerer (Christian.Timmerer@aau.at) Vignesh V Menon Efficient Content-Adaptive Feature-based Shot Detection for HTTP Adaptive Streaming 23