9.Frequently Asked Questions

We would like to inform you about the information in the FAQ for the AI On-site Learning Application.
When using this application, the installation conditions and the method of collecting images for learning will greatly affect the accuracy of detection, so we recommend that you check the site before installation.
For details, please contact your installer or distributor.

9.1About this application in general


No.	Question	Answer
1	Which cameras can use the AI On-site Learning Application?	Please refer to the following URL for the target cameras. →Technical Information <C0103>
2	Can I learn to detect anything if I learn?	The detection performance of the on-site learning function depends on the learning results and the environment in which it is used. It is not suitable for use in use cases where perfect detection performance is required. Please check the operation at the installation site thoroughly before use. We do not guarantee any detection performance.
3	What is suitable or unsuitable for on-site learning?	Objects with clear contours (e.g., special vehicles such as forklifts, certain types of animals, etc.) are suitable for field study. Conversely, objects with indefinite shapes (flames, smoke, rain, etc.) are not suitable.
4	Can you provide a learning model from the beginning that detects various things without learning?	Creating a model that can detect any background and any angle of view requires a large number of images. The advantage of on-site learning is that high detection accuracy can be achieved with a small amount of learning images, since the training can be performed on actual images of the scene to be detected.

9.2About learning


No.	Question	Answer
1	How many images do I need to learn?	Between 10 and 200 images can be used for learning. For objects with little change in reflection, 10 images may be sufficient. If the change in reflection (change in size, change in visibility due to orientation, percentage of background in the frame for learning, color with brightness change) is large, increasing the number of images for learning will make it easier to detect the object.
2	Can I also learn on-site from recorded video?	Yes, you can use them as learning images by pre-converting them to still images in JPEG or PNG format. Please see below for details. →3.2.3　Upload images from PC
3	Can the system learn from images other than those from installed cameras?	Images from other cameras can be used for learning, but it is better to use images from cameras in the field for accuracy.
4	It takes time to set up the bounding box. Is there an easy way to do this?	In case of improving false detection or improving missed detection, there is a function to automatically set detection frames for human, vehicle, and bicycle. For details, please see below. →3.3　Set bounding boxes and learn
5	Do you have any tips for setting up a bounding box?	- Please frame all objects. - Please frame the object so that it contains as little background as possible. Hints are also provided on the frame setting screen for learning.
6	Do I need an SD card to learn?	It is stored in the camera's internal memory, so there is no need for an SD card.
7	When do you use the training data extension?	In real-world environments, when it is not possible to collect images of dark environments or images with opposite orientations, etc., the system can automatically increase the number of variations of learning images. For details, please see below. →4.3　Advanced learning settings
8	When replacing a camera, do I need to relearn it?	By downloading the detector and uploading it to the new camera, you can use the detector as it was before the camera was replaced. For details, please see below. →4.8　Maintenance

9.3About detection


No.	Question	Answer
1	What it has learned is not detected.	Adjust the score threshold. If you have difficulty detecting a particular reflection, add an learning images of that reflection.
2	It will falsely detect things other than what it has learned.	Adjust the score threshold. Adding learning images that show objects prone to false detection may also improve it. The Guidance window also provides tips for learning.
3	What size can be detected?	See below. →1.3　Camera Installation Conditions
4	It is only trained on color images, but can it detect black-and-white images as well?	Although black-and-white images may be detected in some cases, if detection is difficult, adding black-and-white images to the learning or selecting [black-and-white image] in the [Training data extension] may improve the detection.
5	Is it possible to detect objects of different colors than the learned object? Conversely, can only the learned colors be detected?	Although different colors may still be detected, learning with [Do not distinguish between colors] enabled will make it easier to detect different colors. If you want to detect only learned colors, disable the [Do not distinguish between colors] option. For details, please refer to the following. →4.3　Advanced learning settings
6	Can I use the detector with other cameras?	Yes. However, the accuracy of detection in the destination camera depends on how the image is viewed. If the images are similar in reflection to the learned images, detection may be possible; however, to improve detection performance, it is recommended that images taken by each camera be added to the learning.

9.4About On-site learning of forklifts

As a concrete example, this section provides an FAQ on learning a forklift as a new detection object.


No.	Question	Answer
1	Does the entire forklift need to be in the image?	Collect images showing the entire forklift.
2	How do I learn if my forklift is hidden by luggage?	If the area hidden by the cargo is small, attach the bounding box to the area other than the cargo (only the body of the forklift), so that it is less affected by the cargo. If the area hidden by the cargo is large, attach the bounding box to the entire forklift truck, including the cargo.
3	Do they need to learn to forklift forward and backward as well as sideways?	It is necessary to learn the image for the orientation you want to detect.
4	Do I need to learn both forklifts from different manufacturers?	If the shapes are similar, they are likely to be recognizable, but those with different shapes must be learned individually.
5	Is it possible to detect objects of different colors than the learned object? Conversely, can only the learned colors be detected?	Each should be learned as a separate object.
6	Should I have them learn each counter and reach type?	Yes. However, the accuracy of detection in the destination camera depends on how the image is viewed. If the images are similar to the learned images, accurate detection may be possible, but additional learning may be required for each camera to achieve accuracy.