READ ME File For "Efficient Video Recognition with Convolutional Neural Networks by Exploiting Temporal Correlation in Video Data"



ReadMe Author: Amin Sabet, University of Southampton

This dataset supports the PhD thesis:
AUTHORS: Amin Sabet
TITLE: Efficient Video Recognition with Convolutional Neural Networks by Exploiting Temporal Correlation in Video Data



This dataset contains:

Data for Figure 3.1, Figure.3.7 ,Figure.3.10, Figure.3.21 ,Figure.3.13 ,Figure.3.14, Figure.3.15, Figure.3.16, Figure.3.17, Figure.3.18.
Table 4.1, Table 4.2, Table 4.3, Table 4.4


The figures are as follows:

FIGURE 3.1 The averaged of unchanged pixels into all convolutional layers  between  consecutive  frames  when  processing  Video1, video2 and video3.


Figure 3.7 Energy breakdown of RS

Figure 3.10 Comparing  mean  average  precision  of  video  object detection  of  floating-point  precision  (FP32),  INT8,  CBinfer, DeepCach with SQS implementation of YOLOv3.

Figure. 3.12 Comparing the normalized energy consumption for video object detec-
tion of floating-point precision (FP32), CBinfer, DeepCach, and SRS implementations
of YOLOv3 [11]. SRS is repeated for symmetric and asymmetric SQS with γ0 = 0.1,
γ0 = 0.3 γ0 = 0.5

Figure. 3.13  Energy   breakdown   of   Conv2   for   RS   and   SRSdataflows

Figure. 3.14  Energy  Consumption  breakdown  of  Convolutional layers  in  table  3  processed  by  row-stationary  dataflow  (RS)and similarity-aware dataflow (SRS).

Figure. 3.15 Energy consumption and energy breakdown of convolutional layer for different level of similarity between ifmaps.For each convolutional layer, the bars from left to right show the energy breakdown for RS dataflow,

Figure. 3.16 Trade off between energy consumption and quantization error of convolutinal layer

Figure. 3.17 Histogram of similarity between the features of consecutive frames from stationary

Figure. 3.18 Comparing the energy consumption of similarity-aware Yolo and Resnet to that of conventional Yolo and Resnet

Table 4.1 Accuracy, precision, recall, and F1 scores measures for TEEMs in TEE-FR with Resnet101 feature network.

Table 4.2 Performance comparison with the state-of-the-art on the ImageNet VID and TVnet validation set. Our method speed up existing models with negligible mAP error. Run time is measured on X GPU

Table 4.3 Performance comparison of TEEM with conventional feature tracking approaches on a random subset of ImageNetVID and TVnet validation set.

Table 4.4 Computational complexity and speed of TEE-FR video object detection.
 
Licence: CC BY

Related projects: International Centre for Spatial Computation

Date that the file was created: December 2022