A review of Detection/region/object proposal methods

PapersJ. Hosang, R. Benenson, P. Dollár, and B. Schiele. What makes for effective detection proposals? arxiv:1502.05082, 2015.
ArXiv

@ARTICLE {hosang2015arxiv,
  author = {J. Hosang and R. Benenson and P. Doll\ ' ar and B. Schiele},
  title = {What Makes For effective detection proposals.},
  journal = {arxiv:1502.05082}, Year
  = {}}

J. Hosang, R. Benenson, and B. Schiele. How good is detection proposals, really? BMVC 2014.
PDF, ArXiv

@INPROCEEDINGS {HOSANG2014BMVC,
  author = {J. Hosang and R. Benenson and B. Schiele},
  title = {How good is detectio n proposals, really?},
  BookTitle = {BMVC}, Year
  = {}}

Transferred from: http://zhangliliang.com/2015/05/19/paper-note-object-proposal-review-pami15/

Project address of the paper:
https://www.mpi-inf.mpg.de/departments/computer-vision-and-multimodal-computing/research/ object-recognition-and-scene-understanding/how-good-are-detection-proposals-really/ Outline

Depending on the order in which the article is described, the following will probably be: review the various methods of object proposal (hereinafter referred to as OP) and classify them. Discussion on the robustness of different op in the reappearance of the image after disturbance discussion of the different OP Recall on Pascal and Imagenet, where the author presents a new standard for average Recall (AR) Discusses the performance comparisons of the different op for the actual classification (compared with the two well-known detector of DPM and RCNN), and shows that AR is a standard that is quite relevant to performance.

First list of previous effects:

Note that here are just a few ways to find the source code, so let's start with a little bit of sorting. Review of the various OP methods

The author roughly divides the OP method into two categories, called Grouping method, a class called Window scoring method. The former refers to a method that first breaks the picture and then aggregates it, such as selective search. The latter is a way of generating a large number of windows and scoring, and then filtering out the low points, such as objectness. There are also a number of ways between the two, such as Multibox. Grouping Proposal Methods

The author continues to divide the grouping method into three small classes. SP, Superpixel, GC, using seed Point and groupcut to split, EC, extract proposal from edge graph. Here is an introduction to the Selectivesearch (SP): No need to learn, first to break the picture into Superpixel, and then according to the human defined distance aggregation. CPMC (GC): Randomly initialize seed points, then do graphcut to split, repeat multiple times, and then define a very long feature to sort. (So the speed is super slow) MCG (EC): First, using a ready-made method to quickly get a hierarchical segmentation results, and then using the edge information for aggregation. Window Scoring Proposal methods

Unlike the former, which needs to generate candidate boxes by aggregating small blocks, the method here is to make a candidate box, and then direct the scoring sort to filter out the candidate box for the low score. Introducing two well-known methods, Bing: Trained a simple linear classifier to filter candidate boxes in a way like a sliding window, which is surprisingly fast and can reach MS level on the CPU. But in the literature [40] The attack says that classification performance does not come from learning but geometry. Edgeboxes: As with selective search is a method that does not need to learn, combined with sliding window, by calculating the number of the window's inner edge to score, the final sort. aliternate Proposal Methods Multibox, currently the only one based on the CNN extract proposal method, through the CNN to return to the position of the N candidate box and scoring, It should be the first in Imagenet's dectection track. Baseline Proposal Methods

Here with uniform,gaussian,sliding window and superpixels as baseline, not the focus is not to expand said. discussion on the robustness of the various OP methods for reproducibility

The author puts forward the assumption that a good Op method should have good reproducibility, that is, the object retrieved in a similar image should be consistent. The method of verification is to do various disturbances to Pascal's picture (Figure 2), and then see if it can also detect the same object recall is how much, according to the strict IOU can get a curve, and finally calculate the area under the curve repeatability.

Here the chart is a lot more specific please see the original paper, here directly on the author's conclusion, Bing and edgeboxes on the repeatability best performance. the recall of various OP methods

Here the good Op method should have a high recall, or you will miss the detected object. Three ways to measure Recall are discussed here: Recall versus IoU threshold: Fixed proposal quantity, calculated IoU Recall Recall number of versus window according to different proposal criteria S: with 1 complementary, here first fixed IOU, according to the number of different proposal to calculate recall Average recall (AR): The author proposed, here just according to different proposal number, calculate IOU between 0.5 to 1 recall.

data sets, the authors tested the Pascal VOC07 and imagnet Detection datasets.
Here are a lot of pictures, here only one AP, other please refer to the original paper.

or directly on the conclusion MCG, Edgebox,selectivesearch, rigor and geodesic perform well in different proposal numbers if only the proposal,mcg,endres and cpmc of less than 1000 are the best If the position of the candidate box is not well positioned at the beginning, as the IOU standard is strict, the recall will fall faster including Bing, Rahtu, Objectness and edgeboxes. The decline in Bing is particularly noticeable. In AR this standard, MCG performance is stable, endres and edgeboxes in less proposal time performance is better, when allow more proposal time, rigor and selectivesearch performance will be better than others. On Pascal and Imagenet, each Op method is similar, which shows that the performance of these OP methods is good generalization. the effect of various OP methods in actual detection tasks

Here the author of the OP followed by two kinds of detection famous detector to test, one is the literature [54] of Lm-llda (a DPM variant), the other is r-cnn, it is worth noting that the two detector authors are RBG ... The Great God also ...
Here we use various OP methods to extract 1k proposal, and then compare them.
Also directly to the author's conclusion: The more accurate the OP method is, the greater the help for the classifier, because the more accurate the positioning, the higher the score the classifier will return:
Under Lm-llda and R-CNN, the top 5 op methods of the map are Mcg,seletivesearch,edgeboxes,rigor and geodesic.
A list of points is shown below.
Through analysis, the authors found that AR and map have a strong correlation:
The author uses AR as a guide to tuning the parameters of edgeboxes, and then gets a better map (up to 1.7 points)
Summary and discussion of the full text

Summary: For repeatability This standard, the current Op method effect is general. It is possible to improve the repeatablilty of the OP method by more robust features of noise and disturbances. But repeatability low does not mean that the last map is low, such as Selectivesearch, so the end is to see the application scenario. If the OP method is positioned more accurately, the help for the classifier becomes larger. So for the OP method, a IOU of 0.5 recall is not a good standard. High recall but the positioning is not accurate, will hurt the last map Mcg,seletivesearch,edgeboxes,rigor and geodesic is currently the best performance of the 5 methods, the speed of edgeboxes and geodesic for excellent. The current Op method is similar in both VOC07 and imagenet, indicating that they all have good generalization performance.

Discussion: OP is useful if the computational power is up. The authors believe that the sliding window plus the strong classifiers such as CNN will have a better effect if the computational performance permits. The author observes that the features used in the current OP (such as Object boundary and Superpixel) are not used in the classifier, and there are no other op-like features in the OP method other than Multibox. The authors look forward to a job that combines the advantages of both. Finally, the author makes a three-point guess: Then top down may play a more important role in OP, and later the OP and detector will be more closely connected, and the op-generated segmentation mask will play a more important role.