Skip to content

Data collection by temporal consistency of object detection and tracking

License

Notifications You must be signed in to change notification settings

MertCatalkaya/temporal_consistency_odt

 
 

Repository files navigation

Temporal Anomaly Detection in Videos

Data collection through temporal consistency of a video:

This repository harnesses object detection and tracking techniques to identify inconsistencies and anomalies in video sequences. By identifying these discrepancies, we can capture valuable data to enhance our object detection model. These anomalies can be annotated and integrated into the training, validation, and testing phases for future model iterations. Consequently, our tool serves as a core component in an active learning pipeline

The system identifies the following anomalies:

Discrepancies in Object Count:

  • Frames missing objects that are present in adjacent frames.
  • Frames with additional objects not found in adjacent frames.

Class Mismatch:

  • Instances where an object's predicted class differs from its previous frame classification.

Low IOU:

  • Situations where the intersection-over-union (IOU) of an object between two consecutive frames drops below a threshold, indicating potential tracking issues.

Examples

Below are some examples:

- Additional/missing objects

- Different classes

class_i class_j
im1 im2

- Low iou

Below is an example of low_iou (in fact, object_3 is switched to another car). Hence the iou=0.0

frame_i frame_(i+1)
im1 im2

Other outputs

Here is an example where an object is cropped from the video. For the sake of visibility, it is captured with 5 FPS.

Notice that in frame 38, the class name changes from car to truck, and in the next frame it switches back to car.

vid1.mov

Augmentation Evaluation: The repository also supports the application of random frame augmentations. By observing how these augmentations affect detection outcomes, users can assess the resilience and robustness of the detection model (yolo_v8 in our case). This may also aid in accumulating diverse samples that can further improve the model's performance.

Below is an example of random augmentations TRIGGER ALERT: might trigger photosensitive people!!!:

vid2b.mov

Installation

Prerequisite: pyenv

pyenv simplifies Python version management, enabling you to seamlessly switch between Python versions for different project requirements.

https://github.com/pyenv/pyenv-installer

On macOS you can use brew, but you may need to grab the --HEAD version for the latest:

brew install pyenv --HEAD

or

curl https://pyenv.run | bash

And then you should check the local .python-version file or .envrc and install the correct version which will be the basis for the local virtual environment. If the .python-version exists you can run:

pyenv install

This will show a message like this if you already have the right version, and you can just respond with N (No) to cancel the re-install:

pyenv: ~/.pyenv/versions/3.8.6 already exists
continue with installation? (y/N) N

Prerequisite: direnv

direnv streamlines environment variable management, allowing you to isolate project-specific configurations and dependencies within your development environment.

https://direnv.net/docs/installation.html

curl -sfL https://direnv.net/install.sh | bash

Developer Setup

If you are a new developer to this package and need to develop, test, or build -- please run the following to create a developer-ready local Virtual Environment:

direnv allow
python --version
pip install --upgrade pip
pip install poetry
poetry install

References

One of the core pieces of this repo object detection and tracking is obtained from the following article: https://thepythoncode.com/article/real-time-object-tracking-with-yolov8-opencv

About

Data collection by temporal consistency of object detection and tracking

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%