0016E5.MXF 2.54 GB 0006R0.MXF 1.68 GB 0005VD.MXF 1.55 GB 01TP_extract.avi 1.65 GB ------- 7.42 GB Usage: DirectShowWithParser.exe Exameple: > DirectShowWithParser.exe extract_06R0_01Hz_ECCV08seq1.txt Read the individual script files to see what they do exactly. They all extract a series of images that are exactly the images used in our different experiments (ECCV08, PRL08, ICVW08, etc). No lossy compression has taken place after the HD camera's initial encoding. Scripts should all be in scriptsForFrameExtractor.zip. Notes: General ---------- The original footage (included) has more frames than were analyzed in our experiments thus far. The images we colored with semantic labels were sampled from those big sequences at 1Hz. To compute the camera poses, we used 30Hz, and (IMPORTANT, NOT REFLECTED IN THE INCLUDED SCRIPTS) where possible, we extracted 30 frames before the first labeled image and 30 frames after (pre- and post-roll). The codec (COmpressorDECompressor) used here is mostly the Panasonic P2 codec, the one native to our Panasonic HVX200-NTSC camera (note, NOT square pixels). It is freely available from Panasonic, and we trust Panasonic doesn't mind that we're providing the Windows drivers with our data for convenience. Alternate places to find P2 drivers for your OS: http://panasonic.sixbullets.net/ 0016E5.MXF as Seq 2 DayTrain ---------- 1) The range of frame numbers is actually [390, 7920] and [8190, 8640], so after the extract_16E5_01Hz_ECCV08seq2a.txt produces [390, 8640], you can delete the extra frames. 2) Instead of frame #900, #901 was labeled, hence use of extract_16E5_01Hz_ECCV08seq2[ a and b].txt. So after running both scripts, delete frame #900. 0016E5.MXF as CamSeq101 ---------- This sequence of 101 frames of 15 Hz footage was NOT used for the ECCV08 paper. It was used for the Fauqueur et al. 2007 workshop and the 2008 Pattern Recognition Letters papers. Notice that these frames interrupt the "Seq 2 DayTrain" sequence used in ECCV08, so one could easily expand the latter's length. We used this subsequence mostly to test label-propagation. 01TP_extract.avi as Seq 3 DuskTrain and as Seq 5 DuskTest (originally 0001TP.MXF) ---------- Note: If you have the original 01TP.MXF file (over 3GB!), then follow these instructions but use the other scripts that cite 0001TP.MXF. Normally, you should be using 01TP_extract.avi because the relevant data is the same pixel for pixel. This sequence was split, the first half (62 frames) for training, the other for testing. They come from a 3.5Gb MXF file. MXF's are very efficient, but we do not have a way to losslessly cut off the many extra minutes contained in that file, so the decision was made to re-encode just the relevant 3690 frames (+30 frames pre-roll, and 31 frames post-roll) using the lossless LAGS (Lagarith) codec. The codec is free (http://lags.leetcode.net/codec.html) and open-source. Hopefully installing it will actually save time compared to downloading the original big MXF. The script file works the same way, since DirectShow reads MXF's and AVI's, as long as you first install the respective codecs (Panasonic P2 and LAGS).