I would like to train a deep learning framework (TensorFlow) for object detection with a new object category.
As source for the ground truthing I have multiple video files which contain the object (only part of the image contains the object).
How should I ground truth the video? Should I extract frame by frame and label every frame even when those video frames will be quite similar? Or what would be best practise for such a task?
Open source tools are preferred.
It usually works as you described. At lest for the iteration zero:
collect required examples (video)
extract valuable frames from the video (manual or partially automated process)
use OpenCV (or any other tool) to extract required details (bounding box, accurate mask)
assemble a training set
train a model
Here is an example of a training set, produced by the approach described above (see it in action)
For iteration one you might use iteration zero models and significantly improve step 2 and step 3 to increase the training set even more.
I'm trying to solve pretty much the same problem, because it is hard to produce a training set to get accurate segmentation:
(again, here it is in action and other examples)
Basically, start with a semi-manual approach and try to evolve.
Related
Im currenty working on a project at University, where we are using python + tensorflow and keras to train an image object detector, to detect different parts of the root system of Arabidopsis.
Our current ressults are pretty bad, as we do only have about 100 images to train the model with at this moment, but we are currently working on cultuvating more plants in order to get more images(more data) to train the tensorflow model.
We have implemented the following Mask_RCNN model:Github- Mask_RCNN tensorflow
We are looking to detect three object clases: stem, main root and secondary root.
But the model detects main roots incorrectly where the secondary roots are located.
It should be able to detect something like this:Root detection example
Training root data set that we are using right now:training images
What is the usual sample size that is used to train a neural network accurate results?
First off: I think there is no simple rule to estimate the sample size but at least it depends on:
1. Quality of your images
I downloaded the images and I think you need to preprocess them before you can use it to reduce the "problem complexity". In some projects, in which I worked with biological data, a background removal (image - low pass filter) was the key to get better results. But you should definitely remove/crop the area outside the region of your interest (like the tape and the ruler). I would try to get the cleanest data set as possible (including manually adjustments cv2/ gimp/ etc.) to focus the network to solve "the right problem".. After that you could apply some random distortion to make it also work on fuzzy/bad/realistic images as well.
2. The way you work with your data
There are a few tricks that enables you to "expand" your dataset.
Sometimes it's very helpful to let a generator method crop random small patches from your input data. This allows you to work with more batches (on small gpus) and gives your network more "variety", (just think about the conv2d task: if you don't use random cropping your filters will slide over the same areas over and over again (at the same image)). Because of the same reason: apply random distortion, flip and rotate your images.
3. Network architecture
In your case I would prefer a U-Net architecture with a last conv2d output of 3 (your classes) feature maps, a final softmax activation and an categorical_crossentropy, this enables you to play with the depth, because sometimes you need sophisticated architectures to solve a problem (close to 100%) but in your case you just want to see a first working result. So fewer layers and a simple architecture could also help you to get things work. Maybe there are some trained network weights for a U-Net which meets your requirements (search on kaggle for example). Because it is also helpful (to reduce the data you need) to use "transfer learning" -> use the first layers of an network (weights) which is already trained. Using a semantic segmentation the first filters will become something like an edge detection for the most given problems/images.
4. Your mental model of "accurate results"
This is the hardest part.. because it evolves during your project. Eg. in the same moment your networks starts to perform well on preprocessed input images you will start to think about architecture/data changes to make it work on fuzzy images as well. This is why you should start with a feasible problem but always improve your dataset (including rare kinds of roots) and tune your network architecture step by step.
I am new to machine learning field and based on what I have seen on youtube and read on internet I conjectured that it might be possible to count pedestrians in a video using tensorflow's object detection API.
Consequently, I did some research on tensorflow and read documentation about how to install tensorflow and then finally downloaded tensorflow and installed it. Using the sample files provided on github I adapted the code related to object_detection notebook provided here ->https://github.com/tensorflow/models/tree/master/research/object_detection.
I executed the adapted code on the videos that I collected while making changes to visualization_utils.py script so as to report number of objects that cross a defined region of interest on the screen. That is I collected bounding boxes dimensions (left,right,top, bottom) of person class and counted all the detection's that crossed the defined region of interest (imagine a set of two virtual vertical lines on video frame with left and right pixel value and then comparing detected bounding box's left & right values with predefined values). However, when I use this procedure I am missing on lot of pedestrians even though they are detected by the program. That is the program correctly classifies them as persons but sometimes they don't meet the criteria that I defined for counting and as such they are not counted. I want to know if there is a better way of counting unique pedestrians using the code rather than using the simplistic method that I am trying to develop. Is the approach that I am using the right one ? Could there be other better approaches ? Would appreciate any kind of help.
Please go easy on me as I am not a machine learning expert and just a novice.
You are using a pretrained model which is trained to identify people in general. I think you're saying that some people are pedestrians whereas some other people are not pedestrians, for example, someone standing waiting at the light is a pedestrian, but someone standing in their garden behind the street is not a pedestrian.
If I'm right, then you've reached the limitations of what you'll get with this model and you will probably have to train a model yourself to do what you want.
Since you're new to ML building your own dataset and training your own model probably sounds like a tall order, there's a learning curve to be sure. So I'll suggest the easiest way forward. That is, use the object detection model to identify people, then train a new binary classification model (about the easiest model to train) to identify if a particular person is a pedestrian or not (you will create a dataset of images and 1/0 values to identify them as pedestrian or not). I suggest this because a boolean classification model is about as easy a model as you can get and there are dozens of tutorials you can follow. Here's a good one:
https://github.com/aymericdamien/TensorFlow-Examples/blob/master/notebooks/3_NeuralNetworks/neural_network.ipynb
A few things to note when doing this:
When you build your dataset you will want a set of images, at least a few thousand along with the 1/0 classification for each (pedestrian or not pedestrian).
You will get much better results if you start with a model that is pretrained on imagenet than if you train it from scratch (though this might be a reasonable step-2 as it's an extra task). Especially if you only have a few thousand images to train it on.
Since your images will have multiple people in it you have a problem of identifying which person you want the model to classify as a pedestrian or not. There's no one right way to do this necessarily. If you have a yellow box surrounding the person the network may be successful in learning this notation. Another valid approach might be to remove the other people that were detected in the image by deleting them and leaving that area black. Centering on the target person may also be a reasonable approach.
My last bullet-point illustrates a problem with the idea as it's I've proposed it. The best solution would be to alter the object detection network to ouput both a bounding box per-person, and also a pedestrian/non pedestrian classification with it; or to only train the model to identify pedestrians, specifically, in the first place. I mention this as more optimal, but I consider it a more advanced task than my first suggestion, and a more complex dataset to manage. It's probably not the first thing you want to tackle as you learn your way around ML.
I am working on a project that aims to detect objects in certain difficult circumstances. I ran a test with Mask_RCNN on a dataset that contains that specific type of difficult examples and it did a pretty good job in some of them.
But some other examples didn't get detected surprisingly, when there is no obvious reason. To understand the reason behind this performance difference, I've been adviced to use Tensorboard. But then I realized that its mostly used for training phase, as I understood from this video.
At the end of the video, however, they mention about an integration project of Tensorboard, namely the Tensorflow Debugger Integration. But unfortunately I could not find further information regarding the continuation about that feature.
Is there any way to visualize weights and activation maps inside a CNN during inference/evaluation phase?
The main difference between training and inference time for tensorboard will be the global_step value. Most graphs display global step as the x-axis. You can supply your own global step counter if you like, but you'll have to decide what the x-axis should represent to you in this case since "time" isn't really a logical construct during inference. Other tabs such as the images tab don't have a time component, so using them should be the same as during training.
The tensorflow debugger is a nice terminal debugger, but wouldn't really be related to what you're trying to do here. It's certainly not a visualization tool.
Another approach might be to simply generate your own plots and output a set of PDFs with the various visualizations you need using standard tools like matplotlib for each test image. I've found tools like XnView make it really easy to look through a lot of PDF visualizations to understand what's going on. I've used this approach quite effectively. If you want to view many hundreds or thousands of results quickly you might have an easier time if all the visuals are just dumped out to a directory.
When training detection models, are images that are used in real life better (i.e. higher accuracy / mAP) than images of the same object but in the form of stock photo?
The more variety the better. If you train a network on images that all have a white background and expect it to perform under conditions with noisy backgrounds you should expect the results on unseen data to perform worse because the network never had a chance to learn distinguiting features of target object vs. background objects.
If you have images with transparent backgrounds one form of data augmentation that would be expected to improve results would be to place that image against many random backgrounds. The closer you come to realistic renderings of an image the better you can expect your results to be.
The more realistic examples you can augment your training dataset with, the better. Note that it generally does not help to add random noise to your data to generate larger training datasets, it only improves results when your expanded dataset contains realistic variants of the original images in the dataset.
My motto when training neural networks is this: The network will cheat any chance it gets. It will learn impressively well, but given the opportunity, it will take shortcuts. Don't let it take shortcuts. That often translates to: Make the problem harder such that no shortcut exists for it to take. Neural networks often perform better under more difficult conditions because the simplest solution it can arrive at is also the most general purpose. Read up on multi-task learning for some exciting examples that provide great food-for-thought.
I have to analyse some images of drops, taken using a microscope, which may contain some cell. What would be the best thing to do in order to do it?
Every acquisition of images returns around a thousand pictures: every picture contains a drop and I have to determine whether the drop has a cell inside or not. Every acquisition dataset presents with a very different contrast and brightness, and the shape of the cells is slightly different on every setup due to micro variations on the focus of the microscope.
I have tried to create a classification model following the guide "TensorFlow for poets", defining two classes: empty drops and drops containing a cell. Unfortunately the result wasn't successful.
I have also tried to label the cells and giving to an object detection algorithm using DIGITS 5, but it does not detect anything.
I was wondering if these algorithms are designed to recognise more complex object or if I have done something wrong during the setup. Any solution or hint would be helpful!
Thank you!
This is a collage of drops from different samples: the cells are a bit different from every acquisition, due to the different setup and ambient lights
This kind of problem should definitely be possible. I would suggest starting with a cifar 10 convolutional neural network tutorial and customizing it for your problem.
In future posts you should tell us how your training is progressing. Make sure you're outputting the following information every few steps (maybe every 10-100 steps):
Loss/cost function output, you should see your loss decreasing over time.
Classification accuracy on the current batch of your training data
Classification accuracy on a held out test set (if you've implemented test set evaluation, you might implement this second)
There are many, many, many things that can go wrong, from bad learning rates, to preprocessing steps that go awry. Neural networks are very hard to debug, they are very resilient to bugs, making it hard to even know if you have a bug in your code. For that reason make sure you're visualizing everything.
Another very important step to follow is to save the images exactly as you are passing them to tensorflow. You will have them in a matrix form, you can save that matrix form as an image. Do that immediately before you pass the data to tensorflow. Make sure you are giving the network what you expect it to receive. I can't tell you how many times I and others I know have passed garbage into the network unknowingly, assume the worst and prove yourself wrong!
Your next post should look something like this:
I'm training a convolutional neural network in tensorflow
My loss function (sigmoid cross entropy) is decreasing consistently (show us a picture!)
My input images look like this (show us a picture of what you ACTUALLY FEED to the network)
My learning rate and other parameters are A, B, and C
I preprocessed the data by doing M and N
The accuracy the network achieves on training data (and/or test data) is Y
In answering those questions you're likely to solve 10 problems along the way, and we'll help you find the 11th and, with some luck, last one. :)
Good luck!