Hello, I just reviewed Paper 1 again and indeed didn't find any comparison diagrams addressing the "boundary blur" issue. Every mention of "blur" only discusses the problem of internal blur, as seen in the following points: 1. This is a problem of region inconsistency, where the entire truck's area should be labeled as the "truck" category, rather than certain internal regions being labeled as other categories, causing the truck to fragment into several discontinuous parts. This issue is classified under the left-hand restricted receptive field problem, not the boundary blur problem. 2. The bottom half of the image addresses the issues of missing object details and inconsistent results. Missing details refer to small elements like the thin brackets of roadside fences, and inconsistent results refer to small objects of the same category in different images being segmented into different shapes or sizes, all of which actually fall under the region inconsistency problem. 3. Besides this, the paper only provides visual results on the Cityscapes dataset without comparing the results before and after addressing boundary blur. Thank you.
语义边缘检测和语义分割的区别
- 语义边缘检测&语义分割
Semantic Edge Detection vs. Semantic Segmentation- 区别difference
- 任务目标
Task Objective- 语义边缘检测
Semantic Edge Detection
- 识别图像中不同物体之间的边界线或轮廓及语义类别
Identifying the boundaries or contours between different objects in an image and their semantic categories
语义分割
-
将图像中的每个像素分类到特定的语义类别(如人、车、树等) Classify each pixel in the image into a specific semantic category (such as people, cars, trees, etc.)
- 识别图像中不同物体之间的边界线或轮廓及语义类别
- 语义边缘检测
- 输出
output- 语义边缘检测
Semantic Edge Detection- 每个边界线或轮廓像素点的类别标签
Category labels for each boundary line or contour pixel -
- 每个边界线或轮廓像素点的类别标签
- 语义分割
Semantic Segmentation- 每个像素的类别标签
Category labels for each pixel -
- 每个像素的类别标签
- 语义边缘检测
- 技术难点
Technical difficulties- 语义边缘检测
Semantic Edge Detection- 多尺度特征
Multiscale features-
处理对象在不同尺度上的特征,通过组合不同尺度上的信息来判断边缘。 Process the features of objects at different scales and judge edges by combining information at different scales.
-
- 全局理解
Global understanding-
在语义边缘检测中,我们不仅仅是寻找图像中像素值变化的地方(这叫做梯度变化),而是要理解图像中每个物体的边界。例如,在一张有猫的照片中,我们不仅要找到颜色变化的地方,还要知道哪里是猫的边缘,哪里是背景的边缘。 In semantic edge detection, we are not only looking for changes in pixel values in the image (this is called gradient change), but also understanding the boundaries of each object in the image. For example, in a photo with a cat, we not only need to find the areas where the color changes, but also know where the edges of the cat are and where the edges of the background are.
-
- 多任务学习
Multi task learning-
同时进行多个相关任务(如检测边缘和提取纹理),相互增强,提高整体性能。 Simultaneously performing multiple related tasks (such as edge detection and texture extraction) to enhance each other and improve overall performance.
-
- 细粒度识别
Fine grained recognition-
细粒度识别指的是在边缘检测中,不仅仅是大类边界(如整个物体的轮廓),还需要对物体内部的细小部件或细节边界进行识别和区分。例如,在一张汽车的图片中,不仅要识别出汽车的轮廓,还要区分出车门、车轮、车窗等细节的边界。这种细致程度的边界识别称为细粒度识别。 Fine grained recognition refers to edge detection, which not only involves identifying large class boundaries (such as the outline of the entire object), but also needs to recognize and distinguish small components or detailed boundaries inside the object. For example, in a picture of a car, it is not only necessary to recognize the outline of the car, but also to distinguish the boundaries of details such as doors, wheels, and windows. This level of fine-grained boundary recognition is called fine-grained recognition.
-
- 噪声鲁棒性
Noise robustness-
在存在噪声和干扰的情况下,仍能准确检测边缘,保证模型的稳定性和可靠性。 In the presence of noise and interference, it is still possible to accurately detect edges, ensuring the stability and reliability of the model.
-
- 多尺度特征
- 语义分割
Semantic Segmentation- 类别不平衡
Category imbalance-
不同类别的像素数量差别很大,常见类别(例如背景)可能占大多数,而少见类别的数据不足。 The number of pixels in different categories varies greatly, with common categories (such as background) possibly accounting for the majority and rare categories having insufficient data.
-
- 同一区域一致性
Consistency within the same region-
在图像中,同一物体或区域的像素应该被一致地标记为同一类别。例如,在分割一张包含猫的图像时,猫的所有像素都应该被标记为“猫”,而不应有一些像素被错误地标记为“背景”或其他类别。 In an image, pixels of the same object or region should be consistently labeled as belonging to the same category. For example, when segmenting an image containing a cat, all pixels of the cat should be labeled as "cat", and some pixels should not be mistakenly labeled as "background" or other categories.
-
- 类别不平衡
- 语义边缘检测
- 任务目标
- 对比
comparison- 效率考量
Efficiency Considerations-
语义边缘检测:模型在效率上有所优化,只关注边缘区域,处理的像素点较少,相对而言计算可以更高效。
Semantic edge detection: The model has been optimized in terms of efficiency, focusing only on edge regions and processing fewer pixels, making it relatively more computationally efficient.语义分割:需要计算每个像素点的类别标签,通常模型较大,计算量和内存需求相对较大。
Semantic segmentation: It is necessary to calculate the category labels of each pixel, and the model is usually large, with relatively high computational and memory requirements.
-
-
细节处理 Detail handling
-
语义边缘检测:在细节边缘处理上需求更高,对边界的定位精度要求严格。
Semantic edge detection: requires higher precision in detail edge processing and strict accuracy in boundary localization.语义分割:倾向于需要更高精准度的像素级分类,一般对特征的细节抓取不如边缘检测。 Semantic segmentation: tends to require pixel level classification with higher accuracy, and generally does not capture the details of features as well as edge detection.
-
- 实时性
Real time performance-
语义边缘检测:由于处理区域(图像中的高频区域)相对较小,更有可能实现实时性。
Semantic edge detection: Due to the relatively small processing area (high-frequency areas in the image), it is more likely to achieve real-time performance.语义分割:因其复杂性,实时性实现相对较困难。
Semantic segmentation: Due to its complexity, real-time implementation is relatively difficult.
-
- 效率考量
- 关系
relationship-
语义边缘检测可以看作是语义分割的一个子集或者一个步骤,但是类似的方法和论文比较少。
Semantic edge detection can be seen as a subset or step of semantic segmentation, but there are relatively few similar methods and papers.- Towards Unified Depth and Semantic Prediction from a Single Image
-
语义边缘检测在这篇论文中确实是用于语义分割的,它通过提供边缘信息来帮助模型更好地理解图像中的结构和边界,从而提高了分割的质量和精度。
Semantic edge detection is indeed used for semantic segmentation in this paper. It helps the model better understand the structure and boundaries in images by providing edge information, thereby improving the quality and accuracy of segmentation.
-
- Improving Semantic Segmentation via Decoupled Body and Edge Supervision
-
论文提出了一个新范式,通过将图像特征分解为对象的主体部分(body)和边缘部分(edge),来分别处理对象内部的一致性和边界细节。这种方法的核心思想是,良好的语义分割性能需要显式地对对象的主体和边缘进行建模,这对应于图像的低频和高频信息。
The paper proposes a new paradigm that decomposes image features into the main body and edge parts of an object to handle consistency and boundary details within the object, respectively. The core idea of this method is that good semantic segmentation performance requires explicit modeling of the subject and edges of the object, which corresponds to the low-frequency and high-frequency information of the image.
-
语义边缘检测和语义分割通常都是多分类学习
Semantic edge detection and semantic segmentation are usually multi classification learning语义边缘检测和语义分割少数是多标签学习
Semantic edge detection and semantic segmentation are mostly multi label learning-
语义边缘检测:例如肿瘤、血管和器官边界。在这种情况下,一个像素点可能同时属于多个类别的边缘(例如,肿瘤边缘和器官边缘)
Semantic edge detection: such as tumor, vascular, and organ boundaries. In this case, a pixel may belong to multiple categories of edges simultaneously (such as tumor edges and organ edges)语义分割:一个像素点可能同时属于"道路"和"阴影"类别
Semantic segmentation: A pixel may belong to both "road" and "shadow" categories at the same time
比较总结
Comparative summary-
语义边缘检测细节边缘处理和检测精度上更胜一筹。
Semantic edge detection outperforms in detail edge processing and detection accuracy.语义分割整体上的场景理解和准确分类优势明显。
The overall scene understanding and accurate classification advantages of semantic segmentation are obvious.
- Towards Unified Depth and Semantic Prediction from a Single Image
-
- 区别difference
- 性能差异
- 阅读的论文
- 语义边缘检测
- SEAL
-
https://openaccess.thecvf.com/content_ECCV_2018/papers/Zhiding_Yu_SEAL_A_Framework_ECCV_2018_pap er.pdf
-
- RINDNet
- https://arxiv.org/pdf/2108.00616
- DFF
- https://arxiv.org/pdf/1902.09104
- STEAL
- https://arxiv.org/pdf/1904.07934
- SEAL
- 语义分割
- CAT-seg
- https://arxiv.org/pdf/2303.11797
- SAN
- https://arxiv.org/pdf/2302.12242
- FCN
- https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=7478072
- SegNet
- https://arxiv.org/pdf/1511.00561
- ZegFormer
- https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9879044
- CAT-seg
- 语义边缘检测
- performance差别
- 问题:语义边缘检测和语义分割的指标不统一
- 语义边缘检测
- 指标
- optimal dataset scale (ODS)
- Per-Image Best Threshold (OIS)
- Average Precision (AP)
- Intersection over Union (loU)
- Mean F-measure (MF)
- 指标
- 语义分割
- 指标
- 平均交并比(MeanIntersection over Union,mlou)
- Accuracy
- 平均准确率(MeanAccuracy)
- 频率加权交并比(Frequency Weighted Intersection over Union, f.w10)
- 全局准确率(GlobalAccuracy,G)
- 类别平均准确率(ClassAverageAccuracy,C)
- 边界F1分数(Boundary F1-measure,BF)
- 指标
- efficiency差别
- 指标
- 参数数量(Param.)
- GFLOPs (Giga Floating Point Operations per Second)
- 推理时间(Inferencetime)
- 训练时间(Training time)
- 内存消耗(Memory consumption)
- 推理速度(FPS)
- 指标
- 阅读的论文