AiOnIc Acube - AI in your handsSee how an image, a question or a spoken phrase becomes a useful result on AiOnIc. Explore the demonstrations below and the kinds of real-world interfaces they could help build.
Register and tell us about your work and interests. An administrator can review your request and share more in-depth information, including technical details and selected demonstrations.
Register your interests →Turn a captured image into a short description, displayed beside the source image or read aloud.
Choose a new model load for each launch or a warm session for repeated captures.
Generate scene captions from single images or sampled frames in a continuing capture loop.
Models remain resident between captures. The loop processes individual frames; it does not establish continuous temporal understanding.
Locate recognized objects and overlay their bounding boxes, class names and confidence values.
Higher-resolution and faster smaller-input modes are available. Custom products, defects and counting workflows need matching data and evaluation.
Display an estimated gaze direction as an arrow on a detected face.
This is coarse image-coordinate direction, not a calibrated screen point or a measure of attention, intent or medical status.
Ask typed questions and receive generated English replies, with a voice-interface option.
Text chat and its voice extension use the compact model runtime. Tool connections and application knowledge must be supplied separately.
Ask questions in Japanese and receive Japanese text replies, with optional speech input and output.
A small chat model can generate incorrect answers. It is not a verified knowledge base or a live-information service.
Speak a Japanese question, review its recognized text and hear a generated reply.
Speech recognition uses accelerator kernels alongside Japanese chat. Recognition, answer quality and speech synthesis are separate stages; microphone and speaker support depend on the board.
Display predicted robot actions beside recorded observations, including translation, rotation and gripper values.
The demonstration displays replay predictions only; it does not move a robot. Numerical agreement is separate from successful closed-loop manipulation.
Turn USB camera or stored video frames into an edge map. A compact convolutional model extracts fine boundaries, learns to upsample them and fuses the result into a live preview.
Native 320×240 and full 640×480 modes are available. Deployment-specific accuracy must be evaluated with representative images.
Estimate body joints from a USB camera or stored video. Register guide poses in a gallery, associate clips and trigger browser playback after a steady match. Optional USB/UART messages can trigger a connected physical player.
Camera-first input includes stored-video fallback for boards without a webcam. Similarity scores are not calibrated probabilities or a medical assessment.
Track the closest visible face and associate its body joints. Raised arms select command 1, a T-pose selects command 2, and an averaged detected-face count above a chosen threshold selects command 3. Adjustable averaging and lockout reduce repeated triggers.
Camera-first input includes stored-video fallback. USB serial is optional and disabled by default. Face area estimates proximity; detected-face counts may omit people whose faces are hidden. No age or gender classification is performed.
A medical example image becomes a reusable embedding. Select a box to generate and refine a segmentation overlay.
Three supplied examples demonstrate research interaction. This is not a diagnostic system; agreement with ONNX is distinct from clinical accuracy.
Resident depthwise demo: interactive box selection, reusable image embeddings and measured ONNX mask agreement.
Camera or stored video passes through a vision backbone and attention-based object queries to produce labelled bounding boxes.
This hybrid demo is slower than real time. The live preview refreshes independently of detections. Original-model detection agreement is reported separately from dataset accuracy.
Dual-resolution feature branches combine spatial detail and scene context, producing a road-scene class overlay.
The Cityscapes model targets road scenes. Indoor class predictions may be misleading; preview cadence and processed-frame FPS are separate.
Visual demos can capture a live image or sample a stored video. Camera-first mode uses a recording when no camera is detected, so boards without a webcam can still demonstrate supported visual models.
The web interface pairs an image with its caption, detections or predicted actions. Authorized users can review results on a computer or phone that can reach the board service.
Download the required package before going into the field and keep it available offline. Installed models and attached microphones, cameras and speakers vary between boards.
These are demonstration capabilities, not completed deployments. Evaluate representative inputs, supported classes, response time and output quality for the application you intend to build.
Share your profile and areas of interest when registering. Administrator-approved access can include deeper explanations of the model pipeline, conversion methods and selected board demonstrations.
Create an account and share your interests →