published on 2018-12-27
142 views
  • 01:08 1.
    FAST MODE DECISION IN HEVC INTRA PREDICTION, USING REGION WISE CNN FEATURE CLASSIFICATION
  • 00:05 2.
    Slide 2
  • 01:17 3.
    Slide 3
  • 09:08 4.
    Zhang et al. [5] introduced the machine learning based CU depth decision method with a joint SVM classifier.
  • 01:31 5.
    Slide 5
  • 03:46 6.
    Slide 6
  • 00:03 7.
    Slide 7
  • 09:22 8.
    Slide 8
  • 00:12 9.
    Slide 9
  • 01:16 10.
    Slide 10
  • 00:05 11.
    Slide 8
  • 05:29 12.
    Slide 10
  • 00:34 13.
    Slide 11
  • 05:56 14.
    ZFNet
  • 00:02 15.
    ZFNet
  • 05:13 16.
    Slide 13
  • 03:10 17.
    Slide 11
  • 00:04 18.
    Slide 13
  • 00:21 19.
    Slide 14
  • 00:57 20.
    We consider ROI as the rectangular window and defined by a tuple (x, y, h, w), where (x, y) specifies its bottom-left corner and (h, w) its height and width. To generate the region proposals we use a 2×2 sliding window over all positions on ConvNet-5 feat
  • 00:11 21.
    Slide 16
  • 01:36 22.
    Slide 17
  • 01:50 23.
    The above process is repeated for 16 times (4scales × 4 aspect ratios) at a single position and continued for 20×30 feature locations. Overall our prediction process generates 20×30×16 possible anchor boxes.The total number of anchor boxes are then reduce
  • 00:32 24.
    Slide 17
  • 00:12 25.
    Slide 13
  • 09:09 26.
    Slide 17
  • 00:04 27.
    The above process is repeated for 16 times (4scales × 4 aspect ratios) at a single position and continued for 20×30 feature locations. Overall our prediction process generates 20×30×16 possible anchor boxes.The total number of anchor boxes are then reduce
  • 00:04 28.
    The above process is repeated for 16 times (4scales × 4 aspect ratios) at a single position and continued for 20×30 feature locations. Overall our prediction process generates 20×30×16 possible anchor boxes.The total number of anchor boxes are then reduce
  • 00:04 29.
    Slide 13
  • 00:20 30.
    Slide 19
  • 01:36 31.
    Fisher Vector
  • 02:10 32.
    Fisher Vector
  • 10:58 33.
    Our network resulted in 256d local features for FV computation and pooled into a representation with 32 Gaussian components. Finally, it resulted into a 16K-d descriptor. Later we reduced the FV dimension to 1024-d by using the PCA technique. After above
  • 00:02 34.
    Slide 22
  • 01:30 35.
    pi is the predicted probability of anchor ‘i’ being an object. pi* is the ground truth label 1 for +ve anchor or 0 for –ve anchor. Ncls and Nloc are the number of anchors in a mini-batch and the total anchors respectively.‘i’ is the index of an anchor box
  • 00:26 36.
    Slide 17
  • 00:52 37.
    pi is the predicted probability of anchor ‘i’ being an object. pi* is the ground truth label 1 for +ve anchor or 0 for –ve anchor. Ncls and Nloc are the number of anchors in a mini-batch and the total anchors respectively.‘i’ is the index of an anchor box
  • 00:33 38.
    Slide 17
  • 00:14 39.
    pi is the predicted probability of anchor ‘i’ being an object. pi* is the ground truth label 1 for +ve anchor or 0 for –ve anchor. Ncls and Nloc are the number of anchors in a mini-batch and the total anchors respectively.‘i’ is the index of an anchor box
  • 02:35 40.
    Our RPN+CNN network is able to achieve detection accuracy up to 65.28% mPA over the above proposed regions per image.For our model training, a 6 step process is included for the joint optimization and detailed steps are summarized in Table I.
  • 01:33 41.
    Slide 26
  • 00:44 42.
    Slide 27
  • 00:02 43.
    Slide 28
  • 00:15 44.
    Slide 29
  • 00:03 45.
    Slide 30
  • 00:11 46.
    Slide 29
  • 03:06 47.
    Slide 30
  • 01:20 48.
    Slide 31
  • 00:44 49.
    Slide 33
  • 03:56 50.
    Slide 30
  • Index
  • Note
  • Discuss
  • Fullscreen
FAST MODE DECISION IN HEVC INTRA PREDICTION, USING REGION WISE CNN FEATURE CLASSIFICATION
1:37:00, published on 2018-12-27 by VCLab NTHU