{"notes":[{"id":"feature-fusion/hourglass","category":"feature-fusion","legacyPath":"feature-fusion/1603-hourglass","legacySlug":"1603-hourglass","legacyPaths":[],"title":"Hourglass","subtitle":"被遺忘的長者","author":"Z. Yuan","paperDate":"2016-03-22","publishedOn":"2024-04-06","paperTitle":"Stacked Hourglass Networks for Human Pose Estimation","paperUrl":"https://arxiv.org/abs/1603.06937","bibliography":{"title":"Stacked Hourglass Networks for Human Pose Estimation","authors":["Alejandro Newell","Kaiyu Yang","Jia Deng"],"url":"https://arxiv.org/abs/1603.06937","pdfUrl":"https://arxiv.org/pdf/1603.06937","codeUrl":null,"preprint":{"id":"1603.06937","submittedOn":"2016-03-22","url":"https://arxiv.org/abs/1603.06937"},"publication":{"venue":"ECCV","year":2016,"kind":"conference","url":"https://doi.org/10.1007/978-3-319-46484-8_29"},"citation":"@misc{featurefusionhourglass2016,\n  title = {{Stacked Hourglass Networks for Human Pose Estimation}},\n  author = {Alejandro Newell and Kaiyu Yang and Jia Deng},\n  year = {2016},\n  eprint = {1603.06937},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/1603.06937}\n}\n","citationPath":"/content/papers/citations/feature-fusion/hourglass.300cfcc3e581.bib"},"thumbnail":{"src":"/content/papers/assets/ad4d473578cc1a03387cd17083539bba81f4eea08875d248c057e128cd1fce71.jpg","width":1024,"height":468},"locale":"zh-hant","path":"/content/papers/zh-hant/feature-fusion/hourglass.4d36e11a4a8c.md","headings":[{"id":"模型架構","text":"模型架構","depth":2},{"id":"hourglass-的最初","text":"Hourglass 的最初？","depth":2},{"id":"結論","text":"結論","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"feature-fusion/trans-unet","relation":"comparison","topic":"編解碼器如何跨尺度傳遞特徵"},{"id":"feature-fusion/unet","relation":"comparison","topic":"編解碼器如何跨尺度傳遞特徵"}]},{"id":"feature-fusion/trans-unet","category":"feature-fusion","legacyPath":"feature-fusion/2102-trans-unet","legacySlug":"2102-trans-unet","legacyPaths":[],"title":"TransUNet","subtitle":"不能只有卷積","author":"Z. Yuan","paperDate":"2021-02-08","publishedOn":"2025-02-06","paperTitle":"TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation","paperUrl":"https://arxiv.org/abs/2102.04306","bibliography":{"title":"TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation","authors":["Jieneng Chen","Yongyi Lu","Qihang Yu","Xiangde Luo","Ehsan Adeli","Yan Wang","Le Lu","Alan L. Yuille","Yuyin Zhou"],"url":"https://arxiv.org/abs/2102.04306","pdfUrl":"https://arxiv.org/pdf/2102.04306","codeUrl":null,"preprint":{"id":"2102.04306","submittedOn":"2021-02-08","url":"https://arxiv.org/abs/2102.04306"},"publication":null,"citation":"@misc{featurefusiontransunet2021,\n  title = {{TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation}},\n  author = {Jieneng Chen and Yongyi Lu and Qihang Yu and Xiangde Luo and Ehsan Adeli and Yan Wang and Le Lu and Alan L. Yuille and Yuyin Zhou},\n  year = {2021},\n  eprint = {2102.04306},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/2102.04306}\n}\n","citationPath":"/content/papers/citations/feature-fusion/trans-unet.a4dac5dc037b.bib"},"thumbnail":{"src":"/content/papers/assets/14209e3562b4a87f468b64e95c4cdb775196c8c6b7bca43c935875959165e49e.jpg","width":1366,"height":678},"locale":"zh-hant","path":"/content/papers/zh-hant/feature-fusion/trans-unet.1628936c33d9.md","headings":[{"id":"定義問題","text":"定義問題","depth":2},{"id":"解決問題","text":"解決問題","depth":2},{"id":"討論","text":"討論","depth":2},{"id":"結論","text":"結論","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"feature-fusion/unet","relation":"comparison","topic":"編解碼器如何跨尺度傳遞特徵"},{"id":"feature-fusion/hourglass","relation":"comparison","topic":"編解碼器如何跨尺度傳遞特徵"}]},{"id":"feature-fusion/unet","category":"feature-fusion","legacyPath":"feature-fusion/1505-unet","legacySlug":"1505-unet","legacyPaths":[],"title":"U-Net","subtitle":"融合的序幕","author":"Z. Yuan","paperDate":"2015-05-18","publishedOn":"2024-08-15","paperTitle":"U-Net: Convolutional Networks for Biomedical Image Segmentation","paperUrl":"https://arxiv.org/abs/1505.04597","bibliography":{"title":"U-Net: Convolutional Networks for Biomedical Image Segmentation","authors":["Olaf Ronneberger","Philipp Fischer","Thomas Brox"],"url":"https://arxiv.org/abs/1505.04597","pdfUrl":"https://arxiv.org/pdf/1505.04597","codeUrl":null,"preprint":{"id":"1505.04597","submittedOn":"2015-05-18","url":"https://arxiv.org/abs/1505.04597"},"publication":{"venue":"MICCAI","year":2015,"kind":"conference","url":"https://doi.org/10.1007/978-3-319-24574-4_28"},"citation":"@misc{featurefusionunet2015,\n  title = {{U-Net: Convolutional Networks for Biomedical Image Segmentation}},\n  author = {Olaf Ronneberger and Philipp Fischer and Thomas Brox},\n  year = {2015},\n  eprint = {1505.04597},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/1505.04597}\n}\n","citationPath":"/content/papers/citations/feature-fusion/unet.b015405a1c46.bib"},"thumbnail":{"src":"/content/papers/assets/0b302bf8ed079d85b3a07562463380875eb3d9dfc631684549000c38d16a5f38.jpg","width":1224,"height":824},"locale":"zh-hant","path":"/content/papers/zh-hant/feature-fusion/unet.89cc26307d56.md","headings":[{"id":"定義問題","text":"定義問題","depth":2},{"id":"解決問題","text":"解決問題","depth":2},{"id":"討論","text":"討論","depth":2},{"id":"isbi-cell-tracking-challenge-2015","text":"ISBI Cell Tracking Challenge 2015","depth":3},{"id":"結論","text":"結論","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"segmentation/fcn","relation":"background","topic":"密集預測與編解碼器的跳接"},{"id":"feature-fusion/unetpp","relation":"next","topic":"密集預測與編解碼器的跳接"},{"id":"feature-fusion/trans-unet","relation":"comparison","topic":"編解碼器如何跨尺度傳遞特徵"}]}],"text":"如果你是時常在不同論文間徘徊的開發者，Hourglass 架構就算說不上熟悉，也肯定是不陌生的。我們先快速地介紹一下這個架構：\n\nHourglass 架構是專為人體姿勢估計任務而設計。這個架構通過全面處理和整合所有尺度的特徵，致力於最佳地捕捉與人體相關的多種空間關係。研究團隊顯示，通過將自下而上和自上而下的處理方法重複應用，並結合中間的監督訓練，可以極大地提升網路的性能。因此，他們將這一新架構命名為「堆疊沙漏」網路。該網路基於連續的池化和上採樣步驟，用於生成最終的預測結果。\n\n這篇論文提出來的時間比 FPN 還要早了半年多，但為什麼他們的名氣差異如此巨大呢？最直觀的感覺就是論文的引用數，截至 2023.08，FPN 引用數為 20816；Hourglass 引用數為 5365。怎麼都一樣的架構設計，引用數卻可以差距四倍之多？\n\n- **等一下，你說它們有一樣？**\n\n## 模型架構\n\n我們之前已經聊過 FPN 的話題了，它的架構想必你也很熟悉了。而本次要討論的 Hourglass 架構，如下：\n\n![hourglass_1](/content/papers/assets/ad4d473578cc1a03387cd17083539bba81f4eea08875d248c057e128cd1fce71.jpg)\n\n乍看之下，他們是截然不同的架構，但我們實際來順著論文的敘述，把這個模型走過一次：\n\n1. **捕捉不同尺度的信息**\n\n   想像你在觀察一張圖片，上面有一個人的全身。要準確地理解這個人的姿勢，我們需要同時關注整個身體以及細節的部分，比如臉部和手部。但這些信息可能位於不同的尺度上。沙漏設計的目的就是能夠同時捕捉這些不同尺度的信息。\n\n2. **沙漏結構的設計**\n\n   這個模型的結構就像這個沙漏一樣。它由一系列的卷積層（用來提取特徵）和最大池化層（用來減少圖像的解析度）組成。這樣可以在每個解析度下保留空間信息，不讓細節丟失。\n\n3. **上下採樣和特徵組合**\n\n   當網路處理完較低的解析度後，它會開始進行上採樣，就像把畫面放大一樣。同時，它會將不同尺度的特徵組合在一起，這樣就能夠將全身的信息和細節結合起來。\n\n4. **最終預測**\n\n   最後，網路會通過一些處理來生成最終的預測。這個預測是一組熱圖，就像地圖上不同位置的顏色表示不同的特徵。這裡，網路試圖預測在圖像中哪些地方存在著人體的關節，比如手肘、膝蓋等。\n\n最後第四個部分是屬於特定分支領域的應用，先不管他。\n\n我們先把模型架構圖重畫一次：\n\n![hourglass_2](/content/papers/assets/d82447fa497ca9d71e7708093b719b04d9166ffe8ff3b3ef6fd1eabb5332e827.jpg)\n\n綠色框起來的地方，屬於 Backbone 的範疇，他做了 N 個階段的降採樣。論文中有提到這裡的降採樣方式為一連串的卷積層搭配最大池化操作。接著進入上採樣的流程，逐層放大特徵圖的解析度並相加。圖中的 P1～P5 是這篇筆記加上的標籤；原論文的沙漏模組各層維持相同通道數，最後另用兩層 1×1 卷積產生關節熱圖。\n\n看到這邊，應該很明顯了吧？\n\n就以上下採樣與同尺度特徵相加的方式來看，Hourglass 和 FPN 確實很像呀！\n\n當然，他們的應用場景是不一樣的，Hourglass 實際上從圖中最高解析度的那層特徵圖（P1，原論文為 64×64 輸出）產生關節熱圖，並重複堆疊多個沙漏模組、施加中間監督；而 FPN 並沒有特別提到他們堆疊很多層（實務上是可行的），而主要是著重在使用不同解析度的特徵圖（P2～P5；RPN 另加 P6），透過不同感受野（Receptive Field）的特徵圖來達成多尺度物件偵測的目的。\n\n那到底為什麼這兩篇論文的引用數差這麼多呢？\n\n其實我們也只能給一個推測的答案，因為應用場景不一樣，物件偵測比較熱門，這是其一。另外一個可能的原因是因為論文敘事的結構：在 Hourglass 的論文當中，主要強調的是「堆疊很多層」和「中間監督」。\n\n## Hourglass 的最初？\n\n仔細翻閱 Hourglass 的論文，其實可以找到更早的一篇參考文獻，裡面有著非常像的結構：\n\n**[Bottom-Up and Top-Down Reasoning with Hierarchical Rectified Gaussians (2015.07)](https://arxiv.org/abs/1507.05699)**\n\n在這篇文獻中並沒有稱呼該架構為「Hourglass」，他只是說這是一個「Bottom-Up」和「Top-Down」的架構，以下節錄一段論文的原話：\n\n> 本文主要的目的是為了探索了一種「雙向」的結構，同時考慮了自上而下的反饋：神經元受到上層和下層神經元的影響。該文獻將通過將神經元視為具有修正的潛在變數，在一個二次能量函數中進行操作，這可以看作是一個階層修正高斯模型（RGs）的方法。作者展示了 RGs 可以通過一個二次規劃（QP）進行優化，而 QP 則可以通過一個具有修正線性單元的循環神經網路進行優化。這使得 RGs 可以使用 GPU 進行優化的梯度下降來進行訓練。\n\n欸不是，等等，他這是在說什麼？\n\n那試著講簡單一點：\n\n這篇論文探討了一種新的方法，以更貼近修正的方式處理神經元，就像是對它們進行了微調，使其更適應特定任務。這種方法被應用在一個數學模型中，該模型能夠更好地處理圖像等資料。研究人員展示了這個方法可以透過解決一個特殊的數學問題來進行最佳化，而這個問題可以透過一種特殊的循環神經網路來求解。這樣的結構使得我們能夠在運算中更有效地使用硬體資源。\n\n如果還是看不懂也沒關係，反正這不是本文想討論的內容，別忘了我們只是想要找一下最早提出這個架構的文獻而已，下圖就是他們所使用的模型架構：\n\n![hourglass_3](/content/papers/assets/f991a2dc780690258c586d4bdd650a79660949d55066b892cda1030224268298.jpg)\n\n相信看到這個似曾相似的結構，應該也可以認同它與 Hourglass 有關聯；Hourglass 論文也引用了它，但前作的上下路徑與多次迭代會共用權重，不能當成相同的模組。這真的是最早的架構設計了嗎？\n\n實話說，也不一定。\n\n因為比這篇論文再早約兩個月：\n\n- **[U-Net: Convolutional Networks for Biomedical Image Segmentation](https://arxiv.org/abs/1505.04597)**\n\n已經被提出來了，而 U-Net 的多尺度連接的結構其實也是長這樣，差別只在於 U-Net 使用 concat 操作。但可惜的是這篇論文沒有引用 U-Net，因此也沒辦法看到作者對於這篇前作的總結與評價。\n\n## 結論\n\n雖然 Hourglass 和 FPN 是在不同領域中應用的兩種架構。\n\n但，他們在特徵融合網路的本質其實是一樣的，就只是同一種架構的不同用法而已。\n\n在同樣的基礎架構上，為了去解決不同的問題，進而發展出不同的應用流程。其中，FPN 通過特徵金字塔來處理不同尺度的信息，主要用於對象檢測和分割等任務。相比之下，Hourglass 網路則透過多層次的結構在不同細節層次上進行特徵提取，特別適用於密集預測任務，例如姿態估計等任務。\n\n如果你也曾經對這兩個架構的本質感到困惑，那麼希望本篇文章可以讓你得到解答。\n","imageSizes":{"/content/papers/assets/ad4d473578cc1a03387cd17083539bba81f4eea08875d248c057e128cd1fce71.jpg":{"width":1024,"height":468},"/content/papers/assets/d82447fa497ca9d71e7708093b719b04d9166ffe8ff3b3ef6fd1eabb5332e827.jpg":{"width":1024,"height":561},"/content/papers/assets/f991a2dc780690258c586d4bdd650a79660949d55066b892cda1030224268298.jpg":{"width":1024,"height":351}}}
