Presentation Time: 7/11 15:00 - 16:30

Post-Meeting Questions

  1. 介紹傳統影像如何處理 Blur/deblur ? 對於高斯函數image 模糊化與銳利畫以傳統數學方式實作

    1. 展示程式 (已回應於下方)
    2. 展示實際影像 (已回應於下方)

    How do traditional optical techniques deal with blurring or deblurring the images? Please show the hand-on python code for displaying Gaussian function on image smoothing/sharpening.

    1. Show the code (resolved)
    2. Show the change of images under this technique (resolved)

  2. EDVR 釐清

    1. 對影片進行 inference 時的pipeline 流程，以時間軸的方式展開以視覺化(已回應於下方)
    2. 兩個主要function的作用
      1. PCD (Pyramid, Cascading and Deformable Convolution) (已回應於下方)
      2. TSA (Temporal and Spatial Attention) (已回應於下方)

    A better explanation of EDVR technique/algorithm

    1. Expand the video frames as time-lapse style to visualize the flow of how deblurring algorithm is done (resolved)
    2. The purpose of two main functions in EDVR algorithm
      1. PCD (Pyramid, Cascading and Deformable Convolution) (resolved)
      2. TSA (Temporal and Spatial Attention) (resolved)

  3. 投影片修正

    1. 將因為模糊沒有出現 bbox 的那些 frame 畫面獨立出來，取代以影片呈現 (要解決的案例問題)
    2. 加入時間序列的excel/csv (已完成)

    Adjusting the presentation slides

    1. Isolate the frames where bounding boxes didn’t show due to the blur
    2. Create a CSV file and label each frame with the time (resolved)

  4. 深化分析結果

    1. 將模糊程度(blur level)切割成工作區內與工作區外或是分佈
    2. 應用偵測模型，評估效果

    Dig more into the analysis to strengthen the conclusion

    1. Isolate the blur level from entire image to the blur level in the work zone (the area where the object recognition model take effect) vs. the blur level out of the work zone
    2. Validate the assumption of the blur level difference in and out of the work zone by putting them back into recognition model

    Additional Resource: Data Analysis

    F68 Video Deblur Data Analysis

1 (a) Image blur/deblur: a mathematical breakdown with Gaussian Function (traditional way)

Blur (Smoothing)

Gaussian blur equation:

$$

G(x) = \frac{1}{2\sigma^2}e^{-\frac{x^2+y^2}{2{\sigma}^2}} $$

Gaussian blur represented in code:

blur_img = cv2.GassianBlur(img, (0, 0), sigma) #img is image input

Note (for code):

  1. Image input size = output size
  2. kernel size must be odd and positive. If kernel size = 0, then it’s computed from Sigma.
  3. Sigma ≥ 0
  4. If sigmaX and sigmaY (optional) are both 0, then they’re computed from kernel size

Deblur (Sharpening)

$$ G(x) = (1-\alpha)f_0(x) + \alpha f_1(x) + y\\

$$