Related
I'm setting up the new Tensorflow Object Detection API to find small objects in large areas of satellite imagery. It works quite well - it finds all 10 objects I want, but I also get 50-100 false positives [things that look a little like the target object, but aren't].
I'm using the sample config from the 'pets' tutorial, to fine-tune the faster_rcnn_resnet101_coco model they offer. I've started small, with only 100 training examples of my objects (just 1 class). 50 examples in my validation set. Each example is a 200x200 pixel image with a labeled object (~40x40) in the center. I train until my precision & loss curves plateau.
I'm relatively new to using deep learning for object detection. What is the best strategy to increase my precision? e.g. Hard-negative mining? Increase my training dataset size? I've yet to try the most accurate model they offer faster_rcnn_inception_resnet_v2_atrous_coco as i'd like to maintain some speed, but will do so if needed.
Hard-negative mining seems to be a logical step. If you agree, how do I implement it w.r.t setting up the tfrecord file for my training dataset? Let's say I make 200x200 images for each of the 50-100 false positives:
Do I create 'annotation' xml files for each, with no 'object' element?
...or do I label these hard negatives as a second class?
If I then have 100 negatives to 100 positives in my training set - is that a healthy ratio? How many negatives can I include?
I've revisited this topic recently in my work and thought I'd update with my current learnings for any who visit in the future.
The topic appeared on Tensorflow's Models repo issue tracker. SSD allows you to set the ratio of how many negative:postive examples to mine (max_negatives_per_positive: 3), but you can also set a minimum number for images with no postives (min_negatives_per_image: 3). Both of these are defined in the model-ssd-loss config section.
That said, I don't see the same option in Faster-RCNN's model configuration. It's mentioned in the issue that models/research/object_detection/core/balanced_positive_negative_sampler.py contains the code used for Faster-RCNN.
One other option discussed in the issue is creating a second class specifically for lookalikes. During training, the model will attempt to learn class differences which should help serve your purpose.
Lastly, I came across this article on Filter Amplifier Networks (FAN) that may be informative for your work on aerial imagery.
===================================================================
The following paper describes hard negative mining for the same purpose you describe:
Training Region-based Object Detectors with Online Hard Example Mining
In section 3.1 they describe using a foreground and background class:
Background RoIs. A region is labeled background (bg) if its maximum
IoU with ground truth is in the interval [bg lo, 0.5). A lower
threshold of bg lo = 0.1 is used by both FRCN and SPPnet, and is
hypothesized in [14] to crudely approximate hard negative mining; the
assumption is that regions with some overlap with the ground truth are
more likely to be the confusing or hard ones. We show in Section 5.4
that although this heuristic helps convergence and detection accuracy,
it is suboptimal because it ignores some infrequent, but important,
difficult background regions. Our method removes the bg lo threshold.
In fact this paper is referenced and its ideas are used in Tensorflow's object detection losses.py code for hard mining:
class HardExampleMiner(object):
"""Hard example mining for regions in a list of images.
Implements hard example mining to select a subset of regions to be
back-propagated. For each image, selects the regions with highest losses,
subject to the condition that a newly selected region cannot have
an IOU > iou_threshold with any of the previously selected regions.
This can be achieved by re-using a greedy non-maximum suppression algorithm.
A constraint on the number of negatives mined per positive region can also be
enforced.
Reference papers: "Training Region-based Object Detectors with Online
Hard Example Mining" (CVPR 2016) by Srivastava et al., and
"SSD: Single Shot MultiBox Detector" (ECCV 2016) by Liu et al.
"""
Based on your model config file, the HardMinerObject is returned by losses_builder.py in this bit of code:
def build_hard_example_miner(config,
classification_weight,
localization_weight):
"""Builds hard example miner based on the config.
Args:
config: A losses_pb2.HardExampleMiner object.
classification_weight: Classification loss weight.
localization_weight: Localization loss weight.
Returns:
Hard example miner.
"""
loss_type = None
if config.loss_type == losses_pb2.HardExampleMiner.BOTH:
loss_type = 'both'
if config.loss_type == losses_pb2.HardExampleMiner.CLASSIFICATION:
loss_type = 'cls'
if config.loss_type == losses_pb2.HardExampleMiner.LOCALIZATION:
loss_type = 'loc'
max_negatives_per_positive = None
num_hard_examples = None
if config.max_negatives_per_positive > 0:
max_negatives_per_positive = config.max_negatives_per_positive
if config.num_hard_examples > 0:
num_hard_examples = config.num_hard_examples
hard_example_miner = losses.HardExampleMiner(
num_hard_examples=num_hard_examples,
iou_threshold=config.iou_threshold,
loss_type=loss_type,
cls_loss_weight=classification_weight,
loc_loss_weight=localization_weight,
max_negatives_per_positive=max_negatives_per_positive,
min_negatives_per_image=config.min_negatives_per_image)
return hard_example_miner
which is returned by model_builder.py and called by train.py. So basically, it seems to me that simply generating your true positive labels (with a tool like LabelImg or RectLabel) should be enough for the train algorithm to find hard negatives within the same images. The related question gives an excellent walkthrough.
In the event you want to feed in data that has no true positives (i.e. nothing should be classified in the image), just add the negative image to your tfrecord with no bounding boxes.
I think I was passing through the same or close scenario and it's worth it to share with you.
I managed to solve it by passing images without annotations to the trainer.
On my scenario I'm building a project to detect assembly failures from my client's products, at real time.
I successfully achieved very robust results (for production env) by using detection+classification for components that has explicity a negative pattern (e.g. a screw that has screw on/off(just the hole)) and only detection for things that doesn't has the negative pattens (e.g. a tape that can be placed anywhere).
On the system it's mandatory that the user record 2 videos, one containing the positive scenario and another containing the negative (or the n videos, containing n patterns of positive and negative so the algorithm can generalize).
After a while testing I found out that if I register to detected only tape the detector was giving very confident (0.999) false positive detections of tape. It was learning the pattern where the tape was inserted instead of the tape itself. When I had another component (like a screw on it's negative format) I was passing the negative pattern of tape without being explicitly aware of it, so the FPs didn't happen.
So I found out that, in this scenario, I had to necessarily pass the images without tape so it could differentiate between tape and no-tape.
I considered two alternatives to experiment and try to solve this behavior:
Train passing an considerable amount of images that doesn't has any annotation (10% of all my negative samples) along with all images that I have real annotations.
On the images that I don't have annotation I create a dummy annotation with a dummy label so I could force the detector to train with that image (thus learning the no-tape pattern). Later on, when get the dummy predictions, just ignore them.
Concluded that both alternatives worked perfectly on my scenario.
The training loss got a little messy but the predictions work with robustness for my very controlled scenario (the system's camera has its own box and illumination to decrease variables).
I had to make two little modifications for the first alternative to work:
All images that didn't had any annotation I passed a dummy annotation (class=None, xmin/ymin/xmax/ymax=-1)
When generating the tfrecord files I use this information (xmin == -1, in this case) to add an empty list for the sample:
def create_tf_example(group, path, label_map):
with tf.gfile.GFile(os.path.join(path, '{}'.format(group.filename)), 'rb') as fid:
encoded_jpg = fid.read()
encoded_jpg_io = io.BytesIO(encoded_jpg)
image = Image.open(encoded_jpg_io)
width, height = image.size
filename = group.filename.encode('utf8')
image_format = b'jpg'
xmins = []
xmaxs = []
ymins = []
ymaxs = []
classes_text = []
classes = []
for index, row in group.object.iterrows():
if not pd.isnull(row.xmin):
if not row.xmin == -1:
xmins.append(row['xmin'] / width)
xmaxs.append(row['xmax'] / width)
ymins.append(row['ymin'] / height)
ymaxs.append(row['ymax'] / height)
classes_text.append(row['class'].encode('utf8'))
classes.append(label_map[row['class']])
tf_example = tf.train.Example(features=tf.train.Features(feature={
'image/height': dataset_util.int64_feature(height),
'image/width': dataset_util.int64_feature(width),
'image/filename': dataset_util.bytes_feature(filename),
'image/source_id': dataset_util.bytes_feature(filename),
'image/encoded': dataset_util.bytes_feature(encoded_jpg),
'image/format': dataset_util.bytes_feature(image_format),
'image/object/bbox/xmin': dataset_util.float_list_feature(xmins),
'image/object/bbox/xmax': dataset_util.float_list_feature(xmaxs),
'image/object/bbox/ymin': dataset_util.float_list_feature(ymins),
'image/object/bbox/ymax': dataset_util.float_list_feature(ymaxs),
'image/object/class/text': dataset_util.bytes_list_feature(classes_text),
'image/object/class/label': dataset_util.int64_list_feature(classes),
}))
return tf_example
Part of the traning progress:
Currently I'm using tensorflow object detection along with tensorflow==1.15, using faster_rcnn_resnet101_coco.config.
Hope it will solve someone's problem as I didn't found any solution on the internet. I read a lot of people telling that faster_rcnn is not adapted for negative training for FPs reduction but my tests proved the opposite.
I'm trying to define custom loss function for Caffe using Python layer but I can't clarify what is a required output.
Let's a function for the layer is defined as L = sum(F(xi, yi))/batch_size, where L is loss function to be minimized (i.e. top[0]), x is a network output (bottom[0]), y is ground truth label (i.e. bottom[1]) and xi,yi are i-th samples in a batch.
Widely known example with EuclideanLossLayer (https://github.com/BVLC/caffe/blob/master/examples/pycaffe/layers/pyloss.py) shows that backward level in this case must return bottom[0].diff[i] = dL(x,y)/dxi. Another reference I've found shows the same: Implement Bhattacharyya loss function using python layer Caffe
But in other examples I have seen that it should be multiplied by top[0].diff.
1. What is correct? bottom[0][i] = dL/dx or bottom[0].diff[i] = dL/dxi*top[0].diff[i]
Each loss layer may have loss_weight: indicating the "importance" of this specific loss (in case there are several loss layers for the net). Caffe implements this weight as top[0].diff to be multiplied by the gradients.
Let's back off to basic principles: the purpose of back-propagation is to adjust the layer weights according to the ground-truth feedback. The most basic parts of this include "how far off is my current guess" and "how hard should I yank the change lever?" These are formalized as top.diff and learning_rate, respectively.
At a micro level, the ground truth for each layer is that top feedback, so top.diff is the local avatar of "how far off ...". Thus at some point, you need to include top[0].diff as a primary factor in your adjustment computation.
I know this isn't a complete, direct answer -- but I hope it continues to help even after you solve the immediate problem.
I have two questions:
I need to know that is there any difference between the terms Fast Hadamard Transform (FHT) and Fast Walsh-Hadamard Transform (FWHT) ?
Can i use these two terms interchangeably ?
By a Normalized FHT, do we mean dividing all the values of output matrix (after doing FHT on data) by the largest value in the output matrix ?
What is the physical significance of a Normalized Transform ? (for e.g., Normalized FHT)
regards.
According to Wikipedia it looks the same:
https://en.wikipedia.org/wiki/Hadamard_transform
https://en.wikipedia.org/wiki/Walsh_matrix
https://en.wikipedia.org/wiki/Walsh%E2%80%93Hadamard_transform
Up to the normalization factor.
And indeed there's an efficient implementation using the same trick in the FFT algorithm.
I try to implement a people detecting system based on SVM and HOG using OpenCV2.3. But I got stucked.
I came this far:
I can compute HOG values from an image database and then I calculate with LIBSVM the SVM vectors, so I get e.g. 1419 SVM vectors with 3780 values each.
OpenCV just wants one feature vector in the method hog.setSVMDetector(). Therefore I have to calculate one feature vector from my 1419 SVM vectors, that LIBSVM has calculated.
I found one hint, how to calculate this single feature vector: link
“The detecting feature vector at component i (where i is in the range e.g. 0-3779) is built out of the sum of the support vectors at i * the alpha value of that support vector, e.g.
det[i] = sum_j (sv_j[i] * alpha[j]) , where j is the number of the support vector, i
is the number of the components of the support vector.”
According to this, my routine works this way:
I take the first element of my first SVM vector, multiply it with the alpha value and add it with the first element of the second SVM vector that has been multiplied with alpha value, …
But after summing up all 1419 elements I get quite high values:
16.0657, -0.351117, 2.73681, 17.5677, -8.10134,
11.0206, -13.4837, -2.84614, 16.796, 15.0564,
8.19778, -0.7101, 5.25691, -9.53694, 23.9357,
If you compare them, to the default vector in the OpenCV sample peopledetect.cpp (and hog.cpp in the OpenCV source)
0.05359386f, -0.14721455f, -0.05532170f, 0.05077307f,
0.11547081f, -0.04268804f, 0.04635834f, -0.05468199f, 0.08232084f,
0.10424068f, -0.02294518f, 0.01108519f, 0.01378693f, 0.11193510f,
0.01268418f, 0.08528346f, -0.06309239f, 0.13054633f, 0.08100729f,
-0.05209739f, -0.04315529f, 0.09341384f, 0.11035026f, -0.07596218f,
-0.05517511f, -0.04465296f, 0.02947334f, 0.04555536f,
you see, that the default vector values are in the boundaries between –1 and +1, but my values exceed them far.
I think, my single feature vector routine needs some adjustment, any ideas?
Regards,
Christoph
The aggregated vector's values do look high.
I used the loadSVMfromModelFile() located in http://lnx.mangaitalia.net/trainer/main.cpp
I had to remove svinstr.sync(); from the code since it caused losing parts of the lines and getting wrong results.
I don't know much about the rest of the file, I only used this function.
I need a simple and fast way to compare two images for similarity. I.e. I want to get a high value if they contain exactly the same thing but may have some slightly different background and may be moved / resized by a few pixel.
(More concrete, if that matters: The one picture is an icon and the other picture is a subarea of a screenshot and I want to know if that subarea is exactly the icon or not.)
I have OpenCV at hand but I am still not that used to it.
One possibility I thought about so far: Divide both pictures into 10x10 cells and for each of those 100 cells, compare the color histogram. Then I can set some made up threshold value and if the value I get is above that threshold, I assume that they are similar.
I haven't tried it yet how well that works but I guess it would be good enough. The images are already pretty much similar (in my use case), so I can use a pretty high threshold value.
I guess there are dozens of other possible solutions for this which would work more or less (as the task itself is quite simple as I only want to detect similarity if they are really very similar). What would you suggest?
There are a few very related / similar questions about obtaining a signature/fingerprint/hash from an image:
OpenCV / SURF How to generate a image hash / fingerprint / signature out of the descriptors?
Image fingerprint to compare similarity of many images
Near-Duplicate Image Detection
OpenCV: Fingerprint Image and Compare Against Database.
more, more, more, more, more, more, more
Also, I stumbled upon these implementations which have such functions to obtain a fingerprint:
pHash
imgSeek (GitHub repo) (GPL) based on the paper Fast Multiresolution Image Querying
image-match. Very similar to what I was searching for. Similar to pHash, based on An image signature for any kind of image, Goldberg et al. Uses Python and Elasticsearch.
iqdb
ImageHash. supports pHash.
Image Deduplicator (imagededup). Supports CNN, PHash, DHash, WHash, AHash.
Some discussions about perceptual image hashes: here
A bit offtopic: There exists many methods to create audio fingerprints. MusicBrainz, a web-service which provides fingerprint-based lookup for songs, has a good overview in their wiki. They are using AcoustID now. This is for finding exact (or mostly exact) matches. For finding similar matches (or if you only have some snippets or high noise), take a look at Echoprint. A related SO question is here. So it seems like this is solved for audio. All these solutions work quite good.
A somewhat more generic question about fuzzy search in general is here. E.g. there is locality-sensitive hashing and nearest neighbor search.
Can the screenshot or icon be transformed (scaled, rotated, skewed ...)? There are quite a few methods on top of my head that could possibly help you:
Simple euclidean distance as mentioned by #carlosdc (doesn't work with transformed images and you need a threshold).
(Normalized) Cross Correlation - a simple metrics which you can use for comparison of image areas. It's more robust than the simple euclidean distance but doesn't work on transformed images and you will again need a threshold.
Histogram comparison - if you use normalized histograms, this method works well and is not affected by affine transforms. The problem is determining the correct threshold. It is also very sensitive to color changes (brightness, contrast etc.). You can combine it with the previous two.
Detectors of salient points/areas - such as MSER (Maximally Stable Extremal Regions), SURF or SIFT. These are very robust algorithms and they might be too complicated for your simple task. Good thing is that you do not have to have an exact area with only one icon, these detectors are powerful enough to find the right match. A nice evaluation of these methods is in this paper: Local invariant feature detectors: a survey.
Most of these are already implemented in OpenCV - see for example the cvMatchTemplate method (uses histogram matching): http://dasl.mem.drexel.edu/~noahKuntz/openCVTut6.html. The salient point/area detectors are also available - see OpenCV Feature Detection.
I face the same issues recently, to solve this problem(simple and fast algorithm to compare two images) once and for all, I contribute an img_hash module to opencv_contrib, you can find the details from this link.
img_hash module provide six image hash algorithms, quite easy to use.
Codes example
origin lena
blur lena
resize lena
shift lena
#include <opencv2/core.hpp>
#include <opencv2/core/ocl.hpp>
#include <opencv2/highgui.hpp>
#include <opencv2/img_hash.hpp>
#include <opencv2/imgproc.hpp>
#include <iostream>
void compute(cv::Ptr<cv::img_hash::ImgHashBase> algo)
{
auto input = cv::imread("lena.png");
cv::Mat similar_img;
//detect similiar image after blur attack
cv::GaussianBlur(input, similar_img, {7,7}, 2, 2);
cv::imwrite("lena_blur.png", similar_img);
cv::Mat hash_input, hash_similar;
algo->compute(input, hash_input);
algo->compute(similar_img, hash_similar);
std::cout<<"gaussian blur attack : "<<
algo->compare(hash_input, hash_similar)<<std::endl;
//detect similar image after shift attack
similar_img.setTo(0);
input(cv::Rect(0,10, input.cols,input.rows-10)).
copyTo(similar_img(cv::Rect(0,0,input.cols,input.rows-10)));
cv::imwrite("lena_shift.png", similar_img);
algo->compute(similar_img, hash_similar);
std::cout<<"shift attack : "<<
algo->compare(hash_input, hash_similar)<<std::endl;
//detect similar image after resize
cv::resize(input, similar_img, {120, 40});
cv::imwrite("lena_resize.png", similar_img);
algo->compute(similar_img, hash_similar);
std::cout<<"resize attack : "<<
algo->compare(hash_input, hash_similar)<<std::endl;
}
int main()
{
using namespace cv::img_hash;
//disable opencl acceleration may(or may not) boost up speed of img_hash
cv::ocl::setUseOpenCL(false);
//if the value after compare <= 8, that means the images
//very similar to each other
compute(ColorMomentHash::create());
//there are other algorithms you can try out
//every algorithms have their pros and cons
compute(AverageHash::create());
compute(PHash::create());
compute(MarrHildrethHash::create());
compute(RadialVarianceHash::create());
//BlockMeanHash support mode 0 and mode 1, they associate to
//mode 1 and mode 2 of PHash library
compute(BlockMeanHash::create(0));
compute(BlockMeanHash::create(1));
}
In this case, ColorMomentHash give us best result
gaussian blur attack : 0.567521
shift attack : 0.229728
resize attack : 0.229358
Pros and cons of each algorithm
The performance of img_hash is good too
Speed comparison with PHash library(100 images from ukbench)
If you want to know the recommend thresholds for these algorithms, please check this post(http://qtandopencv.blogspot.my/2016/06/introduction-to-image-hash-module-of.html).
If you are interesting about how do I measure the performance of img_hash modules(include speed and different attacks), please check this link(http://qtandopencv.blogspot.my/2016/06/speed-up-image-hashing-of-opencvimghash.html).
Does the screenshot contain only the icon? If so, the L2 distance of the two images might suffice. If the L2 distance doesn't work, the next step is to try something simple and well established, like: Lucas-Kanade. Which I'm sure is available in OpenCV.
If you want to get an index about the similarity of the two pictures, I suggest you from the metrics the SSIM index. It is more consistent with the human eye. Here is an article about it: Structural Similarity Index
It is implemented in OpenCV too, and it can be accelerated with GPU: OpenCV SSIM with GPU
If you can be sure to have precise alignment of your template (the icon) to the testing region, then any old sum of pixel differences will work.
If the alignment is only going to be a tiny bit off, then you can low-pass both images with cv::GaussianBlur before finding the sum of pixel differences.
If the quality of the alignment is potentially poor then I would recommend either a Histogram of Oriented Gradients or one of OpenCV's convenient keypoint detection/descriptor algorithms (such as SIFT or SURF).
If for matching identical images - code for L2 distance
// Compare two images by getting the L2 error (square-root of sum of squared error).
double getSimilarity( const Mat A, const Mat B ) {
if ( A.rows > 0 && A.rows == B.rows && A.cols > 0 && A.cols == B.cols ) {
// Calculate the L2 relative error between images.
double errorL2 = norm( A, B, CV_L2 );
// Convert to a reasonable scale, since L2 error is summed across all pixels of the image.
double similarity = errorL2 / (double)( A.rows * A.cols );
return similarity;
}
else {
//Images have a different size
return 100000000.0; // Return a bad value
}
Fast. But not robust to changes in lighting/viewpoint etc.
Source
If you want to compare image for similarity,I suggest you to used OpenCV. In OpenCV, there are few feature matching and template matching. For feature matching, there are SURF, SIFT, FAST and so on detector. You can use this to detect, describe and then match the image. After that, you can use the specific index to find number of match between the two images.
Hu invariant moments is very powerful tool to compare two images
Hash functions are used in the undouble library to detect (near-)identical images (disclaimer: I am also the author). This is a simple and fast way to compare two or more images for similarity. It works using a multi-step process of pre-processing the images (grayscaling, normalizing, and scaling), computing the image hash, and the grouping of images based on a threshold value.