Use opencv to detect faces using 11 different caffe models

Although many tutorials on using opencv and cv2.dnn.readnetfromcaffe exist over the internet , all of them just use the same model and don't reveal the fact that not all models work the same way.
I will be sharing some data related to the performance of  all of those models.
You can also check my tutorial on caffe models and how to read them.
I, as many others, use opencv mainly for image augmentation or processing.
I used opencv for real time applications on embedded systems such as raspberry pi or nvidia jetson.
Here i will be showing the results for two aspect ratios 1.0, 1.0/255.0 .
The results shown will be the ones with the class human, but i will be showing the overall results as well with a confidence less than the max reported confidence for an object minus 0.3 as i have seen some weird results.
Some models work with an aspect ratio and produce completely wrong results on the other.
Some models detect a different set of faces on each aspect ratio.
I tested the model on some images chosen from the fddb dataset to measure the model's accuracy on hardly recognized faces, most of the faces are occluded in some way, and it has front faces, side faces, statues, drawings and reflections of faces on glass.

VGG VOC0712 SSD 300x300 iter 160000
VGG is a neural network for detection that is trained on pascal's visual object classes's dataset for the years 2007 and 2012 challenges.
The network is trained as a single shot multibox detector with input images of size 300x300.
Its trained for 160000 iterations.
It took 0.5s on average to detect faces.
It detected about 52% of the faces keeping in mind 50% is not bad at all when considering how hard the dataset is.

                         VGG_VOC0712 SSD 512x512 iter 160000
VGG network is trained on VOC07 and VOC12 datasets.
Its trained as a single shot detector for input images of size 512x512.
Its trained for 160000 iterations.
It took 1.2s on average to detect faces.

It detected about 43% of the faces as well but they were many times different than those detected by the 300x300 model.




               VGG ILSVRC2016 SSD 300x300 iter 440000
VGG network is trained on imagenet's large scale visual recognition challenge for the 2016 competition.
Its trained as a single shot detector with size 300x300.
its trained for 440000 iterations.
It took 0.7s on average to detect faces.
Although it detected more than 48% of the faces, it detected some false faces and it detected each face multiple times, solving that problem can be as easy as making the confidence threshold higher or combining bboxes with more than 70% intersection. 
Its quite robust to different sizes, rotations and quality problems but its not as good as the others with occlusions.




                                 MobileNet VOC0712 SSD
MobileNet network is trained on MS-COCO dataset then fine tuned on pascal's visual object classes's dataset for the years 2007 and 2012 challenges.
The one trained on MS-COCO first had a classification mean average precision of 0.727 .
The one not trained on MS-COCO had a maP of 0.68 as reported by the developers.
It took 0.05s on average to detect faces.
It detected 48% of the faces.

                        VGG VOC0712 SSD 512x512 iter 120000
VGG network is trained on VOC07 and VOC12.
Its trained as a single shot detector for input images of size 512x512.
Its trained for 120000 iterations.
It took 1.2s on average to detect faces.
It detected 46% of the faces.


                            VGG COCO SSD 512x512 
VGG network is trained on the 2007 coco dataset only.
Its trained as a single shot detector for input images of size 512x512.
It took 1.2s on average to detect faces.
It detected 46% of the faces.

                             VGG COCO SSD 300x300
VGG network is trained on the 2007 coco dataset only.
Its trained as a single shot detector for input images of size 512x512.
It took 0.4s on average to detect faces.
It detected 47% of the faces.


                                   VGG VOC0712 SSD 300x300 iter 120000
VGG network is trained on VOC07 and VOC12 datasets.
Its trained as a single shot detector for input images of size 300x300.
It took 0.5s on average to detect faces.



                    VGG_VOC0712 SSD 512x512 iter 240000
VGG network is trained on VOC07 and VOC12 datasets.
Its trained as a single shot detector for input images of size 512x512.
Its trained for 160000 iterations.
It took 1.3s on average to detect faces.

Comments

Post a Comment

Popular posts from this blog

Create a route optimization algorithm with zero costs using google's OR-tools and OSRM Part 3

This is a continuation to the route optimization series . First we define the function that will be used by the optimizer to get the distances between points.   def distance_callback(from_index, to_index):                 from_node = manager.IndexToNode(from_index)                  to_node      = manager.IndexToNode(to_index)                          if from_node == to_node:                                return 0                          return data['distance_matrix'][from_node][to_node] IndexToNode is used to correlate between a location as a node index inside the being tested/optimized route by the solver and the index of the loc...
Read more

Learn python programming through algorithms - Binpacking part 2

Image
Bin packing algorithm for regular and irregular shapes Part 2: Constructing the Binary tree (linked list) s is the size, it is a tuple containing the length and the width. c is the coordinate, it is a tuple that will contain its origin point and the origin point won't be the center. Origin point for an inserted shape in red The pointers right and left are used to construct the tree that we will discuss later. occupants is used to hold some complex shapes such that the sorting function takes care of splitting the Node while sorting, thats one method to defeat stuttering as we defer some of the splitting overhead in the inserting stage to the sorting stage. class Node:            def __init__(self, a=[0,0], b=[0,0]):                    self.s = a                    self.c = b                ...
Read more

Create a route optimization algorithm with zero costs using google's OR-tools and OSRM Part 1

Image
This is a tutorial for an algorithm that has riped over the years of experience in doing algorithms and websites  for carpooling, Uber like service, Freight routing, and many other applications. I will be stating the current algorithm and i will discuss my decisions on certain code implementations as well as some of the previous decisions that i took and proved to be wrong, slow or error prone. This algorithm is capable of optimizing for thousands of locations, and made to be highly customizable. Features of the algorithm are: Reading data from and exporting data to an excel sheet or a Database. Multiple time windows per location. Multiple vehicle types, capacities, buffers, depots. Custom optimization parameters like distance, time, fuel consumption, etc. Pickup and dropoff. Custom loading and unloading time per stop. Custom service time at certain locations. Custom max idle time waiting for the next stop's time window. Custom distributed workload amongst ...
Read more