Academia.eduAcademia.edu

Figure 3 – uploaded by Basura Fernando

See full PDFdownloadDownload figure
Figure 3: Vector value representations for VideoDarwin . For a random video, we see the signal for (a) the original in- dependent frames, (b) moving average and (c) time varying mean vectors. Each colour represents a dimension. In (d), (e) and (f) y axis shows the predicted ranking score of each frame obtained from signal (a), (b) and (c) respectively after applying the ranking function (prediction ranking value at t =u! .y;).

Figure 3 Vector value representations for VideoDarwin . For a random video, we see the signal for (a) the original in- dependent frames, (b) moving average and (c) time varying mean vectors. Each colour represents a dimension. In (d), (e) and (f) y axis shows the predicted ranking score of each frame obtained from signal (a), (b) and (c) respectively after applying the ranking function (prediction ranking value at t =u! .y;).

Related Figures (11)

Figure 1: Illustration of how VideoDarwin works. In this video, as Emma moved out from the house, the appearance of the frames evolves with time. A ranking machine learns this evolution of the appearance over time and returns a ranking function. We use the parameters of this ranking function as a new video representation which captures vital information about the action. firstname. lastname@esat.kuleuven.be
Figure 1: Illustration of how VideoDarwin works. In this video, as Emma moved out from the house, the appearance of the frames evolves with time. A ranking machine learns this evolution of the appearance over time and returns a ranking function. We use the parameters of this ranking function as a new video representation which captures vital information about the action. firstname. lastname@esat.kuleuven.be
Figure 2: Processing steps of VideoDarwin for action recognition. First, we extract frames x1... x, from each video. Then we generate feature vz for frame t by processing frames from x, to x¢ as explained in section 3.2. Afterwards, using ranking machines we learn the video representation u for each video. Finally, video specific u vectors are used as a representation for action classification.
Figure 2: Processing steps of VideoDarwin for action recognition. First, we extract frames x1... x, from each video. Then we generate feature vz for frame t by processing frames from x, to x¢ as explained in section 3.2. Afterwards, using ranking machines we learn the video representation u for each video. Finally, video specific u vectors are used as a representation for action classification.
Table 1: Comparison of different video representations for VideoDarwin . Results reported in mAP on the Holywood2 dataset using FDVD with Fisher vectors. As also motivated in Sec. 3.2, the time varying mean vector representation captures better the video-wide temporal information present in a video.
Table 1: Comparison of different video representations for VideoDarwin . Results reported in mAP on the Holywood2 dataset using FDVD with Fisher vectors. As also motivated in Sec. 3.2, the time varying mean vector representation captures better the video-wide temporal information present in a video.
Table 2: One-vs-all accuracy on HMDBS1 dataset [12]
Table 2: One-vs-all accuracy on HMDBS1 dataset [12]
Table 3: Results in mAP on Hollywood? dataset [19]
Table 3: Results in mAP on Hollywood? dataset [19]
Table 4: Results in mAP on MPII Cooking fine grained ac- tion dataset [25].
Table 4: Results in mAP on MPII Cooking fine grained ac- tion dataset [25].
Figure 4: Per class AP in the Hollywood? dataset. The AP is improved for all classes significantly, with an exception of “Drive car’, where context already provides useful infor- mation.
Figure 4: Per class AP in the Hollywood? dataset. The AP is improved for all classes significantly, with an exception of “Drive car’, where context already provides useful infor- mation.
Figure 5: Mean class similarity obtained with (left) max- pooling and (right) VideoDarwin on MPII Cooking activi- ties dataset using BOW-based MBH features extracted on dense trajectories. Non-linear forward VideoDarwin are used for our method. Table 5: Detailed analysis of precision and recall on the ChaLearn gesture recognition dataset [2]
Figure 5: Mean class similarity obtained with (left) max- pooling and (right) VideoDarwin on MPII Cooking activi- ties dataset using BOW-based MBH features extracted on dense trajectories. Non-linear forward VideoDarwin are used for our method. Table 5: Detailed analysis of precision and recall on the ChaLearn gesture recognition dataset [2]
Table 7: Comparison of the proposed approach with the state-of-the-art methods on ChaLearn gesture recognition dataset sorted by reverse chronological order. Table 6: Comparison of the proposed approach with the state-of-the-art methods sorted by reverse chronological or- der. Results reported in mAP for Hollywood2 and Cooking datasets. For HMDBS51 we report one-vs-all classification accuracy.
Table 7: Comparison of the proposed approach with the state-of-the-art methods on ChaLearn gesture recognition dataset sorted by reverse chronological order. Table 6: Comparison of the proposed approach with the state-of-the-art methods sorted by reverse chronological or- der. Results reported in mAP for Hollywood2 and Cooking datasets. For HMDBS51 we report one-vs-all classification accuracy.
Figure 6: Comparison of action recognition performance after removing some frames from each video randomly on Hollywood2. VideoDarwin appears to be stable even when up to 20% of the frames are missing.
Figure 6: Comparison of action recognition performance after removing some frames from each video randomly on Hollywood2. VideoDarwin appears to be stable even when up to 20% of the frames are missing.
Related topics:
Action Recognition
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved