Hello, Hello, everyone. I love the world, the little flower blooming Domingo.
Many are addicted sister, [my talent, but I still can not believe handsome, ha ha]
Body: continuous update. . . Stay tuned
Video whether pumping must draw a frame, or randomly selected frames per second, each video or fixed are fixed by 30, that in the end how much influence? ?
1- pumping a second, to give CAP FPS, the fixed interval, the same length when the last frame number, such as 1min, it is 60;
2- randomly selected, randomly selected 30 pure, Random generates a random number index;
3- 30 are fixed by, regardless of how much fps video, each video frame interval may not be evacuated, if a total of 600, then the break 20
According to the third method currently extracted features inceptionV3 , see if the following effect.
3.1 * 30 2048 First averaging, to obtain dimensional feature 2048, but also by the other embodiment wherein the negative
3.2 2048 by PCA dimensionality reduction to 1000
3.3 The above data obtained via t-SNE 2-dimensional or 3 dimensional data and visualization
1 worry: when extracted by the feature inceptionV3 picture no special treatment, directly resize the 299, whether this effect feature extraction effect? ?
Or join point torch in Compose for good? That is, a variety of crop methods
The following dimensions are reduced to the PCA 800, t-SNE perplexity as a result of 50, the effect is not good feeling.
