Learning pseudo labels for semi-and-weakly supervised semantic segmentation
Pattern Recognition2022
Authors: Yude Wang, Jie Zhang, Meina Kan, Shiguang Shan
Introduction
Semi-and-weakly supervised semantic segmentation簡稱SWSSS,意旨所有data都有label,差在於label的程度不同。
Image Not Showing
Possible Reasons
- The image was uploaded to a note which you don't have access to
- The note which the image was originally uploaded to has been deleted
Learn More →
每個pixel都帶有label的資訊,稱為strong labeled data
Image Not Showing
Possible Reasons
- The image was uploaded to a note which you don't have access to
- The note which the image was originally uploaded to has been deleted
Learn More →
每張圖只帶有image class資訊,稱為weakly labeled data
作者認為現今SWSSS領域有三種路線:
- 簡單將少量strong labeled data與大量weak labeled data結合起來訓練,在訓練中pixel-level supervision的方式佔主導地位,最後導致模型嚴重偏向strong labeled data(overfitting)
- 直接採取weakly supervised semantic segmentation方式,基於CAM作image-level的學習,但由於CAM生成的pseudo label邊框較不準確,可能導致模型因錯誤的pseudo label參與訓練造成效能下降。
- 著重在設計模型以防止錯誤的pseudo label影響訓練,然而這種模組的設計方面往往受到許多限制。
作者將semi-and-weakly supervised semantic segmentation分解成兩部分,一塊是pseudo label的生成,一塊是模型的再訓練,本篇論文針對pseudo label的生成提出有效的改動方法。
Contribution
- 作者改動原本的cross entropy提出class-aware cross entropy(CCE)的loss function,可以更有效區分同時出現的class分佈
- 作者提出PCT方式可動態計算預測分數並挑選好的模型作為teacher給另一個模型作supervised訓練
Class-aware cross entropy (CCE)
Image Not Showing
Possible Reasons
- The image was uploaded to a note which you don't have access to
- The note which the image was originally uploaded to has been deleted
Learn More →
傳統的cross entropy需仰賴大量pixel label才可有效區分每個class,不符合SWSSS的只有少量pixel-level label情況,且因pixel-level label本身可以無負擔的轉換成image-level label,作者提出利用每個data本身帶有的image-level label去過濾掉不存在在圖中的class,如左半邊的分佈圖所示,此方式可有效區分同時出現的class分佈。
Image Not Showing
Possible Reasons
- The image was uploaded to a note which you don't have access to
- The note which the image was originally uploaded to has been deleted
Learn More →
若能使用image-level label資訊過濾掉不存在途中的class,則可更容易幫助模型作pixel-level label的預測,例如以上圖例對於狗的毛皮若有多種class則難以分辨是甚麼動物的毛,若能過濾到只剩背景及狗兩種類別則可讓模型預測更準確。
Image Not Showing
Possible Reasons
- The image was uploaded to a note which you don't have access to
- The note which the image was originally uploaded to has been deleted
Learn More →
f(.)代表network;X = f(Image)
Image Not Showing
Possible Reasons
- The image was uploaded to a note which you don't have access to
- The note which the image was originally uploaded to has been deleted
Learn More →
以上式子為傳統的cross entropy作pixel-wise的計算,利用下列式子:
Image Not Showing
Possible Reasons
- The image was uploaded to a note which you don't have access to
- The note which the image was originally uploaded to has been deleted
Learn More →
將原先pixel-level label轉為image-level label形式,並加入算式中
Image Not Showing
Possible Reasons
- The image was uploaded to a note which you don't have access to
- The note which the image was originally uploaded to has been deleted
Learn More →
若不存在的class則其預測結果不會被列入計算
Binary cross entropy for image-level labeled data
Image Not Showing
Possible Reasons
- The image was uploaded to a note which you don't have access to
- The note which the image was originally uploaded to has been deleted
Learn More →
對image-level label的訓練採用Binary cross entropy的方式:
Image Not Showing
Possible Reasons
- The image was uploaded to a note which you don't have access to
- The note which the image was originally uploaded to has been deleted
Learn More →
Progressive cross training method(PCT)
Image Not Showing
Possible Reasons
- The image was uploaded to a note which you don't have access to
- The note which the image was originally uploaded to has been deleted
Learn More →
Hybrid-Net採用交雜訓練方式,同時訓練pixel-level及image-level兩種lebel data。
Hybrid-to-Pseudo (H2P) training:
Image Not Showing
Possible Reasons
- The image was uploaded to a note which you don't have access to
- The note which the image was originally uploaded to has been deleted
Learn More →
Pseudo-Net利用Hybrid-Net生成的pseudo label做pixel-level的訓練
Pseudo-to-Hybrid (P2H) training:
Hybrid-Net反過來利用Pseudo-Net生成的pseudo label訓練,利用Pseudo-Net的預測結果評估信賴分數作為權重,控制P2H對訓練過程的影響。
Image Not Showing
Possible Reasons
- The image was uploaded to a note which you don't have access to
- The note which the image was originally uploaded to has been deleted
Learn More →
Ji是一般mIOU的計算方式,在每次訓練過程持續做更新
Image Not Showing
Possible Reasons
- The image was uploaded to a note which you don't have access to
- The note which the image was originally uploaded to has been deleted
Learn More →
若class不存在此圖中則不會被列入考慮,最終選取存在圖中最低的mIOU作為控制權重
Framework
Image Not Showing
Possible Reasons
- The image was uploaded to a note which you don't have access to
- The note which the image was originally uploaded to has been deleted
Learn More →
Image Not Showing
Possible Reasons
- The image was uploaded to a note which you don't have access to
- The note which the image was originally uploaded to has been deleted
Learn More →