Notes on "Single Shot MultiBox Detector"

Notes on Single Shot MultiBox Detector by Liu et al (2016):

 This paper introduces Single Shot MultiBox Detector (SSD) which is a feedforward convolutional neural network that prodcues a fixed size collection of bounding boxes and scores for the instances of those bounding boxes, followed by a non minimal suppression step to produce the final detections. Compared to YOLO this algorithm has following features:
  1. Multi-scale feature maps: Unlike YOLO that operates with a single feature map, multi-sacle feature maps are used; convolution layers for detection are added with decreasing size which allows dectecting at multiple scales.
  2. Convolutional predictors for detection: Extra feature layers produce predictions for detection; Each feature layers (at fixed sizes) produce either the score predictions or the offsets in bounding boxes (absolute box positions relative to each feature map positions).
  3. Default boxes and aspect ratios: A set of default bounding boxes are associated with each feature map.
The implementation details on training SSD are as follows:
  1. Matching strategy: Using the Jacard overlap coefficient the ground truth box and the sets of default boxes are compared; all the boxes that are over 0.5 are taken to simplify the learning problem.
  2. Training objective: The overall objective is a weighted sum of the localization (Smooth L1 loss with offsets to the center, and the width and height of the bounding box) and the confidence loss (softmax over multiple classes confidences). 
  3. Scales and aspect ratio for the boxes: Rules are presented for choosing the sets of default boxes in terms of scales and aspect ratio (it is done evenly spaced way). Optimal selection is still an open question.
  4. Hard negative mining: Because many sets of default boxes are tried out, there exists imbalance between negatives and positives for confidences (large negatives will exist). By picking the negative default boxes using confidences the ratio of 3:1 between negatives and positives are kept which was found to make the training more stable.
  5. Data augmentation: To increase the robustness with respect to the size and the shapes, the data is augmented by randomly patching the images.
Note that pre-trained VGG-16 on the ILSVRC CLS-LOC dataset is used as the base neural network in this paper. Details on the training (choce of hyper parameters and so on) are mentioned in the paper and the open-source codes do exist.

 The data sets PASCAL VOC2007, PASCAL VOC2012 and COCO are used to compare SSD to various algorithms. SSD is found to be, atleast in the experiment set up the author used, competitive in both accuracy and speed to other state-of-the-art object detection algorithms.

 Find the overview below.


Q&A:

 Q1) Different overlap coefficients?
A) https://stats.stackexchange.com/questions/238684/what-are-the-difference-between-dice-jaccard-and-overlap-coefficients

Comments

Post a Comment

Popular posts from this blog

Notes on "Focal Loss for Dense Object Detection" (RetinaNet)

Image
Focal Loss for Dense Object Detection by Lin et al (2017) The central idea of this paper is a proposal for a new loss function to train one-stage detectors which works effectively for class imbalance problems (typically found in one-stage detectors such as SSD). The authors also introduce RetinaNet based on the proposed loss function which outperforms the state-of-the-art detection algorithms including the two-stage detectors. Focal loss function: pt is defined as p if y = 1 and p-1 otherwise. In this way, CE(pt) is the cross entropy where p is the probability score. As shown below, the cross entropy is denoted blue (gamma = 0). By adding the factor (see the focal loss FL(pt)) the loss function computes smaller value for the well classified examples and puts more focus on hard and misclassified eamples. In this way the easy background examples are dealt with, which enables the training of highly accurate desne object detection algorithms. Just like alpha-cross entropy is the comm
Read more

Introduction to Bayesian methods

Image
Introduction to Bayesian methods: 1. Bayesian way of thinking: A man is running. Why? Some of the possible explanations are (1) this man is in hurry, (2) this man is running, (3) this man just spotted a dragon, and (4) this man always runs. There are three principles in Bayesian way of thinking which we can use to conclude which of these explanations are the most reasonable. These principles are: 1. Use the prior knowledge on the problem - there is no dragon. 2. Choose the answer that explains the data the most - the man is not wearing the runner. 3. Avoid making extra assumptions - crude to think that any man always run. The answer number "1. this man is in hurry" should be chosen following this principles. 2. Review of probability: Checklists are: 1. What is the probability? 2. What are random variables? 3. What is the probability mass function (PMF)? 4. What is the probability density function (PDF)? 5. What does it mean for the random variables X and Y
Read more

Conjugate Priors

Conjugate Priors. 1. Analytical inference: See the Bayes formulae. The problem is that the evidence is difficult to estimate for the analytical inference. One way to get away is the maximum posterior (MAP) and is very easy to compute, but it comes with many problems. 1. Lack of the in-variance to re-parameterization. 2. The MAP cannot be used as the prior. 3. The MAP is an untypical point (the mode). 4. The credible regions may not be computed. 2. Conjugate distributions. Another approach to avoid computing the evidence is called the conjugate distributions. As the likelihood and the evidence is fixed by our model, we can only vary the prior so that it is easier to compute the posterior. I f the posterior distribution p(θ|x) are in the same family as the prior distribution p(θ), the prior and posterior are then called conjugate distributions, and the prior is called a conjugate prior for the likelihood. For example, the Gaussian family is conjugate to itself (or self-
Read more