Paul Browne Selected MSc Work

Slide1: 

Segmenting Digital Video Paul Browne Centre for Digital Video Processing Dublin City University

Slide2: 

Presentation Introduction Evaluation Baseline Shot Boundary Detection Methods Evaluation of Methods Combining Shot Boundary Algorithms Combining Results Scene Segmentation Luminance Scene Segmentation Results Television News Segmentation Television News Segmentation Results Conclusions 2#23

Slide3: 

Digital video is composed of : Frame Shots Shot boundary Scenes Audio Introduction Navigating digital video. Segmentation needed to replace 3#23

Slide4: 

Introduction Shot Segmentation problems Examples object motion person moves into a camera shot ... camera motion panning, zooming … lighting changes camera flash , lightning .. some types of shot boundary dissolves , fades ... digital effects swirls , morphing … To reduce false shot changes Threshold values - higher values Empirical restrictions - example: shot must be greater than 100 frames ….. 4#23

Slide5: 

Evaluation Baseline Eight hours of continuous digital video. Recorded 12th June 1998 from 1pm to 9pm. Broken up into 24 * 20 minute video segments. Manually evaluated for shot boundary changes. Results outputted to baseline log file. For each segment: Video file audio file baseline log file Baseline segmented into 13 programs 5#23

Slide6: 

Shot Boundary Detection Methods Colour Histogram Colour percentages for a frame is stored. Results compared with that of the adjacent frame. Difference value calculated. Difference above a certain value (threshold) is shot change Histogram values generated difference value above certain value is shot change Compared with previous frames histogram values 6#23

Slide7: 

Shot Boundary Detection Methods Edge Detection frame turned into a grayscale image. edge detection algorithm is then applied to the image. difference value calculated for two adjacent frames. difference above a certain value (threshold) is shot change difference value Compared with previous frames edge values 7#23

Slide8: 

Macroblock works on compressed MPEG digital video. Frame split into fixed regions called macroblocks Three types of macroblock I encoded independently of other macroblocks P encode not the region but the motion vector and error block of the previous frame B same as above except that the motion vector and error block are encoded from the previous or next frame Detecting shot changes specific numbers of macroblock types will occur i i i i i i i i P i B B B Frame with macroblocks P P Shot Boundary Detection Methods 8#23

Slide9: 

Evaluation of Methods Two evaluation measures are: Number of correct shots found Number of correct shots found Actual number of shots Number of correct shots found + false shots Recall : Precision : 9#23 There is a balance between these measures

Slide10: 

Evaluation of Methods The following Venn diagram shows the overlap in correct shot boundaries detected by each of the methods. 241 419 23 52 391 281 4449 Colour Histogram Edge Detection Macroblock 10#23

Slide11: 

Evaluation of Methods Average Precision values over 8 hours Colour Histogram 90.4 Edge Detection 90.0 Macroblock 87.4 Average Recall values over 8 hours Colour Histogram 78.9 Edge Detection 70.2 Macroblock 75.3 Programs with lowest Recall values are: Home & Away (Australian soap) Cooking Program 11#23

Slide12: 

Combining Shot Boundary Algorithms Points of note Limit of 356 extra shots by combining Highest Precision and Recall using Colour Histogram Combining Combining favors histogram Problem Additional false changes will also be introduced 12#23

Slide13: 

Combining Shot Boundary Algorithms Logic of the combining method that selects a shot boundary: if difference value(s) above threshold value(s) then shot boundary Colour Histogram Edge Detection Macroblock or or Method(s) difference value Thresholds Shot boundary Low Histogram 13#23

Slide14: 

Combining Results Colour histogram method (best performing method) Precision average on 8 hours : 90.4 % Recall average on 8 hours : 78.9 % Combined method: Precision decreased an average : 1% or 37 shots Recall increased an average : 4% or 167 shots 14#23

Slide15: 

Scene Segmentation Two Approaches Luminance based segmentation Television News Segmentation Problems Scene is a semantic concept Computer needs wide domain knowledge Typical scene will contain many large changes in light and colour over its duration 15#23

Slide16: 

Luminance Scene Segmentation Method designed to detect location based scenes Method operation: Compare adjacent shots using existing shot boundary results Look for large changes in light to detect scene changes Those above threshold are selected as candidates When all shots compared apply a second low threshold to all candidate scenes Finally apply a minimum gap between scenes 16#23

Slide17: 

Luminance Scene Segmentation Results Results are reasonable for situation comedies Algorithm will not segment action programs well Nt # of scenes. Nf # full scene boundaries detected Ns # of valid scenes groups found Ni # false scene groups Baseline content Nt Nf Ns Ni Keeping up Appearances 10 2 15 2 Shortland Street 9 3 6 3 Fair City 13 3 12 4 Shortland Street 17 5 13 6 17#23

Slide18: 

Luminance Scene Segmentation Results False scene detection 18#23

Slide19: 

Television News Segmentation News is highly structured content There is a structure for News scenes that is generally followed 1. Scene begins in studio, introduced by newscaster 2. Move to location shot, view of newsperson or voice at scene 3. Video segment(s) of the main topic is shown Algorithm three step process to detect scenes: 1. Obtain shots with a length of 280 frames (11.2 seconds) 2. Generate colour histogram for candidate shots and their adjacent shot. Remove those with only a small difference value. 3. Generate a comparison of the 20th frame with the 100th frame in the shot. The comparison is an average of the pixel difference of the two frames. 19#23

Slide20: 

Television News Segmentation Results Results for this algorithm are good compared to the luminance approach Nt # of news anchor persons Ns # of news anchors detected by the algorithm. Ni # of false news anchors. Nd # of missing news anchors. Program Nt Ns Ni Nd Details RTE1 9pm news 10 9 3 1 24th October 2000 RTE1 9pm news 6 6 0 3 15th September 2000 RTE1 9pm news 16 14 3 2 4th October 2000 RTE1 1pm news 11 9 2 2 29th November 2000 TV3 7pm news 16 13 3 6 4th December 2000 TV3 7pm news 13 9 4 4 5th December 2000 20#23

Slide21: 

Television News Segmentation Results Correctly Identified scenes Incorrectly Identified scene 21#23

Slide22: 

Conclusions Separating the baseline into logical programs gives a better view of how the shot boundary detection method performs. It is possible to improve the overall Recall performance of shot boundary methods by combining them. Precision and Recall performance will depend on the threshold levels used There is a trade-off between Precision and Recall Future video indexing, retrieval and summarisation may require a higher performance than any single shot boundary is able to deliver. Scene segmentation is feasible on highly structured content like news and location based scenes Automatic scene segmentation will introduce additional errors 22#23

Slide23: 

The End For more information pbrowne@compapp.dcu.ie Center for Digital Video Processing Dublin City University : http://lorca.compapp.dcu.ie/Video/ 23#23