I am comparing models for the detection of objects for maritime Search and Rescue (SAR) purposes. From the models that I used, I got the best results for the improved version of YOLOv3 for small object detection and for FASTER RCNN.
For YOLOv3 I got the best mAP#50, but for FASTER RCNN I got better all other metrics (precision, recall, F1 score). Now I am wondering how to read it and which model is really better in this case?
I would like to add that there are only two classes in the dataset: small and large objects. We chose this solution because the objects' distinction between classes is not as important to us as the detection of any human origin object.
However, small objects don't mean small GT bounding boxes. These are objects that actually have a small area - less than 2 square meters (e.g. people, buoys). Large objects are objects with a larger area (boats, ships, canoes, etc.).
Here are the results per category:
And two sample images from the dataset (with YOLOv3 detections):
The mAP for object detection is the average of the AP calculated for all the classes. mAP#0.5 means that it is the mAP calculated at IOU threshold 0.5.
The general definition for the Average Precision(AP) is finding the area under the precision-recall curve.
The process of plotting the model's precision and recall as a function of the model’s confidence threshold is the precision recall curve.
Precision measures how accurate is your predictions. i.e. the percentage of your predictions that are correct. Recall measures how good you find all the positives. F1 score is HM (Harmonic Mean) of precision and recall.
To answer your questions now.
How to read it and which model is really better in this case?
The mAP is a good measure of the sensitivity of the neural network. So good mAP indicates a model that's stable and consistent across difference confidence thresholds. In your case faster rcnn results indicate that precision-recall curve metric is bad compared to that of Yolov3, which means that either faster rcnn has very bad recall at higher confidence thresholds or very bad precision at lower confidence threshold compared to that of Yolov3 (especially for small objects).
Precision, Recall and F1 score are computed for given confidence threshold. I'm assuming you're running the model with default confidence threshold (could be 0.25). So higher Precision, Recall and F1 score of faster rcnn indicate that at that confidence threshold it's better in terms of all the 3 metric compared to that of Yolov3.
What metric should be more important?
In general to analyse better performing model, I would suggest you to use validation set (data set that is used to tune hyper-parameters) and test set (data set that is used to assess the performance of a fully-trained model).
Note: FP - False Positive FN - False Negative
On validation set:
Use mAP to select best performing model (model that is more stable and consistent) out of all the trained weights across iterations/epochs. Use mAP to understand whether model should be trained/tuned further or not.
Check class level AP values to ensure model is stable and good across the classes.
As per use-case/application, if you're completely tolerant to FNs and highly intolerant to FPs then to train/tune the model accordingly use Precision.
As per use-case/application, if you're completely tolerant to FPs and highly intolerant to FNs then to train/tune the model accordingly use Recall.
On test set:
If you're neutral towards FPs and FNs, then use F1 score to evaluate best performing model.
If FPs are not acceptable to you (without caring much about FNs) pick the model with higher Precision
If FNs are not acceptable to you (without caring much about FPs) pick the model with higher Recall
Once you decide metric you should be using, try out multiple confidence thresholds (say for example - 0.25, 0.35 and 0.5) for given model to understand for which confidence threshold value the metric you selected works in your favour and also to understand acceptable trade off ranges (say you want Precision of at least 80% and some decent Recall). Once confidence threshold is decided, you use it across different models to find out best performing model.
Related
I'd like to know how the training performance changes over the course of the training. Is there any way to access that via Vertex AI automl service?
Unfortunately it is not possible to see the training performance over the course of the training. Vertex AI Auto ML only shows if the training job is running or not.
The only available information is "how well did model performed with the test set after training". This can seen in the "Evaluation" tab in AutoML. You can refer to Vertex AI Auto ML Evaluation for further readings.
AutoML provides evaluation metrics that could help you determine the performance of your model. Some of the evaluation metrics are precision, recall, and confidence thresholds. These vary depending on what AutoML product you are using.
For example if you have an image classification model, the following are the available evaluation metrics:
AuPRC: The area under the precision-recall (PR) curve, also referred to as average precision. This value ranges from zero to one, where a
higher value indicates a higher-quality model.
Log loss: The cross-entropy between the model predictions and the target values. This ranges from zero to infinity, where a lower value
indicates a higher-quality model.
Confidence threshold: A confidence score that determines which predictions to return. A model returns predictions that are at this
value or higher. A higher confidence threshold increases precision but
lowers recall. Vertex AI returns confidence metrics at different
threshold values to show how the threshold affects precision and
recall.
Recall: The fraction of predictions with this class that the model correctly predicted. Also called true positive rate. Precision: The
fraction of classification predictions produced by the model that were
correct.
Confusion matrix: A confusion matrix shows how often a model correctly predicted a result. For incorrectly predicted results, the
matrix shows what the model predicted instead. The confusion matrix
helps you understand where your model is "confusing" two results.
I have used to below hyper parameters to train the model.
rcf.set_hyperparameters(
num_samples_per_tree=200,
num_trees=250,
feature_dim=1,
eval_metrics =["accuracy", "precision_recall_fscore"])
is there any best way to choose the num_samples_per_tree and num_trees parameters.
what are the best numbers for both num_samples_per_tree and num_trees.
There are natural interpretations for these two hyper-parameters that can help you determine good starting approximations for HPO:
num_samples_per_tree -- the reciprocal of this value approximates the density of anomalies in your data set/stream. For example, if you set this to 200 then the assumption is that approximately 0.5% of the data is anomalous. Try exploring your dataset to make an educated estimate.
num_trees -- the more trees in your RCF model the less noise in scores. That is, if more trees are reporting that the input inference point is an anomaly then the point is much more likely to be an anomaly than if few trees suggest so.
The total number of points sampled from the input dataset is equal to num_samples_per_tree * num_trees. You should make sure that the input training set is at least this size.
(Disclosure - I helped create SageMaker Random Cut Forest)
I have been using sklearn to learn on some data. This is a binary classifcation task and I am using a RBF kernel. My data set is quite unbalanced (80:20) and I'm using only 120 samples, with 10ish features (I've been experimenting with a few less). Since I set class_weight="auto" the accuracy I've calculated from a cross validated (10 folds) gridsearch has dropped dramatically. Why??
I will include a couple of validation accuracy heatmaps to demonstrate the difference.
NOTE: top heatmap is before classweight was changed to auto.
Accuracy is not the best metrics to use when dealing with unbalanced dataset. Let's say you have 99 positive examples and 1 negative example, and if you predict all outputs to be positive, still you will get 99% accuracy, whereas you have mis-classified the only negative example. You might have gotten high accuracy in the first case because your predictions will be on the side which has high number of samples.
When you do class weight = auto, it takes the imbalance into consideration and hence, your predictions might have moved towards center, you can cross-check it using plotting the histograms of predictions.
My suggestion is, don't use accuracy as performance metric, use something like F1 Score or AUC.
100 periods have been collected from a 3 dimensional periodic signal. The wavelength slightly varies. The noise of the wavelength follows Gaussian distribution with zero mean. A good estimate of the wavelength is known, that is not an issue here. The noise of the amplitude may not be Gaussian and may be contaminated with outliers.
How can I compute a single period that approximates 'best' all of the collected 100 periods?
Time-series, ARMA, ARIMA, Kalman Filter, autoregression and autocorrelation seem to be keywords here.
UPDATE 1: I have no idea how time-series models work. Are they prepared for varying wavelengths? Can they handle non-smooth true signals? If a time-series model is fitted, can I compute a 'best estimate' for a single period? How?
UPDATE 2: A related question is this. Speed is not an issue in my case. Processing is done off-line, after all periods have been collected.
Origin of the problem: I am measuring acceleration during human steps at 200 Hz. After that I am trying to double integrate the data to get the vertical displacement of the center of gravity. Of course the noise introduces a HUGE error when you integrate twice. I would like to exploit periodicity to reduce this noise. Here is a crude graph of the actual data (y: acceleration in g, x: time in second) of 6 steps corresponding to 3 periods (1 left and 1 right step is a period):
My interest is now purely theoretical, as http://jap.physiology.org/content/39/1/174.abstract gives a pretty good recipe what to do.
We have used wavelets for noise suppression with similar signal measured from cows during walking.
I'm don't think the noise is so much of a problem here and the biggest peaks represent actual changes in the acceleration during walking.
I suppose that the angle of the leg and thus accelerometer changes during your experiment and you need to account for that in order to calculate the distance i.e you need to know what is the orientation of the accelerometer in each time step. See e.g this technical note for one to account for angle.
If you need get accurate measures of the position the best solution would be to get an accelerometer with a magnetometer, which also measures orientation. Something like this should work: http://www.sparkfun.com/products/10321.
EDIT: I have looked into this a bit more in the last few days because a similar project is in my to do list as well... We have not used gyros in the past, but we are doing so in the next project.
The inaccuracy in the positioning doesn't come from the white noise, but from the inaccuracy and drift of the gyro. And the error then accumulates very quickly due to the double integration. Intersense has a product called Navshoe, that addresses this problem by zeroing the error after each step (see this paper). And this is a good introduction to inertial navigation.
Periodic signal without noise has the following property:
f(a) = f(a+k), where k is the wavelength.
Next bit of information that is needed is that your signal is composed of separate samples. Every bit of information you've collected are based on samples, which are values of f() function. From 100 samples, you can get the mean value:
1/n * sum(s_i), where i is in range [0..n-1] and n = 100.
This needs to be done for every dimension of your data. If you use 3d data, it will be applied 3 times. Result would be (x,y,z) points. You can find value of s_i from the periodic signal equation simply by doing
s_i(a).x = f(a+k*i).x
s_i(a).y = f(a+k*i).y
s_i(a).z = f(a+k*i).z
If the wavelength is not accurate, this will give you additional source of error or you'll need to adjust it to match the real wavelength of each period. Since
k*i = k+k+...+k
if the wavelength varies, you'll need to use
k_1+k_2+k_3+...+k_i
instead of k*i.
Unfortunately with errors in wavelength, there will be big problems keeping this k_1..k_i chain in sync with the actual data. You'd actually need to know how to regognize the starting position of each period from your actual data. Possibly need to mark them by hand.
Now, all the mean values you calculated would be functions like this:
m(a) :: R->(x,y,z)
Now this is a curve in 3d space. More complex error models will be left as an excersize for the reader.
If you have a copy of Curve Fitting Toolbox, localized regression might be a good choice.
Curve Fitting Toolbox supports both lowess and loess localized regression models for curve and curve fitting.
There is an option for robust localized regression
The following blog post shows how to use cross validation to estimate an optimzal spaning parameter for a localized regression model, as well as techniques to estimate confidence intervals using a bootstrap.
http://blogs.mathworks.com/loren/2011/01/13/data-driven-fitting/
I'm doing some work on tracking moving objects using a ceiling mounted downward facing camera. I've got to the point where I can detect the position of the desired object in each frame.
I'm looking into using a Kalman filter to track the object's position and speed through the scene and I've reached a stumbling block. I've set up my system and have all the required parts of the Kalman filter except the measurement variance.
I want to be able to assign a meaningful variance to each measurement to allow the correction phase to use the new information in a sensible manner. I have several measures assigned to my detected objects which could in theory be useful in determining how accurate the position should be and it seems logical to try and combine them to derive a suitable variance.
Am I approaching this in the right manner and if so, can anyone point me in the right direction to continue?
Any help greatly appreciated.
I think you are right. According to this post:
Sensor fusioning with Kalman filter
determining the variance is 100% experimental. It seems to me you have everything you need to get good estimates of the variance.
sorry for the late reply. I have personally encountered the same problem in my previous project. I found the advice given by Gustaf Hendeby in his Sensor Fusion lecture slides ( Page 10 of the slides) extremely valuable.
To summarize:
(1) The SNR of your measurement noise and your process noise determines your filter behavior. A high process noise/measurement noise ration makes your filter slower (low-pass filter), which will usually allow smoother tracking, vice versa a if you set your measurement noise low, you essentially have a high pass filter, which tends to have more jitter.
(2) There are numerous papers in the literature discuss on how to set these noise model properly. However, usually a lot of "tuning" is needed depends on your application. Usually the measurement noise is what we can measure/characterize based on the hardware specification. Therefore a recommendation is to fix "R" (measurement noise covariance) and tune Q (the process model noise covariance).