Object detection in images and videos is a widely studied and practiced in computer vision. As the need for automation in image datasets is growing, it demands that the objects present in images are accurately detected. The performance of these detections solely depends on the quality of the dataset and annotations.
For detection tasks, images need to be annotated to indicate the location of class and the region of interest within the image. The annotation is fast and easy to be performed however the traditional techniques used axis aligned bounding boxes that are not fit for objects being aligned vertically (a person) or horizontally (a car) and the four points do not describe the context of the object in the image properly.
In order to encounter this issue, in this blog, we will look at the advantages of rotating bounding boxes over traditional axis aligned bounding boxes. We will also cover the industrial use cases of rotating bounding boxes and finally we will look at the Keymakr’s annotation tools to accurately annotate your data for custom problems.
Benefits of rotating bounding boxes over traditional techniques
For detection images need to be accurately annotated to indicate the region of interest and the class number thus rotating bounding boxes have many advantages over traditional techniques, some of them are:
- Better localization: Traditional, axis aligned bounding boxes do not localize elongated and rotated objects well due to scale and random rotations of objects within the box on the other hand rotating bounding boxes annotations provide a tighter fit for elongated objects so it performs better in remote sensing images.
- Reduced overlap: Rotating bounding boxes reduce overlap in densely packed objects in an image because of their better localization whereas in traditional, axis aligned bounding boxes there is a higher potential overlap because of the way they are localized.
- Isolation of objects: Axis aligned bounding boxes fails to isolate objects in crowded images as adjacent objects overlap, inclusion of background or adjacent features degrades isolation but rotating bounding boxes capture true aspect ratio of objects and contain fewer background pixels.
- Automatic and iterative retraining: Using pretrained model based on rotating bounding boxes annotations with region of interest transformers, it is possible to generate automatic annotation proposals from a similar dataset or from the same dataset and using vertical retraining it is possible to reannotate the objects and automatically retrain it to generate automatic annotations for the problem at hand on the other hand since axis aligned bounding boxes has noise within the box the automatic proposal method may produce incorrect results at times.
Industrial uses of rotating bounding boxes
Accurate object annotations helps the computer vision models to understand the distinct features of the objects within the images to recognize and localize different classes of object and that is the reason this state of the art latest technique is used in various industries that include:
Rotating bounding boxes are being used in the agriculture industry that mainly includes the wheat industry and is used to predict the size and density of the wheat heads that enables the farmers to make informed decisions on wheat growth and cultivation time thus reducing manual efforts immensely.
Experiments have also shown that the traditional axis aligned bounding boxes annotations have failed to fit the objects properly within images with many elongated wheat heads because of rotational variance.
Remote sensing for military applications:
Rotating bounding boxes are also currently being used in military applications that mainly includes military ship tracking, tank monitoring, military air fleet counting and target tracking using aerial satellite imagery, this enables the military to elevate efficiency of their missions and provides security from intruders.
Experiments have shown that axis aligned bounding boxes produces boxes overlap or fail to fit boxes at time making rotating bounding boxes superior in this case too.
Daily many thefts are reported from car parking spaces and thus rotating bounding boxes are currently used to maintain count of vehicles, similarly they are also being used to identify suspicious items such as weapons and brief cases because of their better localization ability and less noise and pocessing the property of rotational invariance.
Livestock monitoring is extremely important as animals often behave differently in the presence of humans. Moreover, the presence of humans can be a source of stress for the animals and can lead to changes in behavior.
It also allows for early recognition of diseases and animal welfare issues more importantly it also enables us to monitor potential wild animals attack so that the danger can be warded off quickly therefore rotating bounding boxes are a better fit in classifying livestock and identifying animal attacks.
Rotating bounding boxes are also currently being used for automated fishery management, cargo management and counting, vessel traffic service and naval warfare as they are able to accurately capture the rotational and translation property of the object within the boxes and this provide higher accuracy under complex occlusion.
Annotation service advantages
Rotating bounding boxes can make the model’s performance highly accurate. Keymakr is a video and image annotation service provider that offers unique advantages for AI innovators.
- In-house team: Our in-house team of annotators is led by experienced managers. In-house teams can provide a higher level of service and quality compared to crowdsourced annotation or remote workers.
- Annotation platform: Keymakr’s proprietary annotation platform is designed to accelerate annotation. A full suite of annotation tools and unique project management capabilities make it easier to meet deadlines for training datasets.
- Verification: Keymakr employs three layers of human verification in addition to an automated quality control check.