👽 Sipeed MaixCAM

Edge computer vision at the source. Run local vision models on a Sipeed MaixCAM and send the resulting structured detection events to your private Companion Memory. Raw video and camera frames stay on the device.

Overview

Companion Capture for MaixCAM is a MaixPy v4  app for the Sipeed MaixCAM and MaixCAM Pro. It runs the models you enable on the camera’s live feed, logs every detection as a JSON event in a per-session file on the SD card, and, when you configure it, uploads each finished session to your Companion Memory capture ingest API.

It’s a capture surface, like the Companion desktop capture app. It isn’t a Companion Hub app: you install it on the camera, not from the Hub app store, and the camera talks to Companion Memory directly with an API key.

Key features

• On-device inference — models run on the MaixCAM’s NPU with the maix runtime. No raw video leaves the device.
• 13 selectable models — 8 computer vision, 3 audio, and 2 media entries, chosen on a touch setup screen. See Models.
• Session logs — each capture session writes one JSON file under /sd/logs/ with a session ID, start and end times, and every detection event.
• Sync to Companion Memory — optional. When enabled, each ended session’s events upload to /v1/events. Uploads wait while Wi-Fi is down and retry with backoff. Nothing is deleted locally until a batch is confirmed uploaded.
• Thermal throttling — AI inference is skipped for a frame when the CPU is at or above the configured temperature (75 °C by default).

Supported hardware

Both run MaixPy v4 (MaixCDK-based) firmware. You also need an SD card for the configuration file and session logs. For the hardware, see the Sipeed MaixCAM documentation .

Models

The setup screen groups models as follows. Every model you enable loads into the same small memory pool at startup, so a device with two or three models on runs faster than one with all of them.

On a fresh install these are on: Object Classification, Face Recognition, Face Emotion Detection, Audio Recording, Audio Playback, Real-time Voice Recognition, Video Recording, and Video Playback.

The app expects these model files, in MaixPy’s .mud format, at these default paths. A missing file disables only that model.

Object Tracking and Counting needs no model file. Hand Gesture Classification falls back to a geometric check on the hand landmarks when its model file doesn’t load. Real-time Voice Recognition needs a keyword-spotting model and a keyword table that the app doesn’t supply, so it does nothing until you provide both.

Architecture

Inference happens on the camera. Only JSON detection events cross the network.

flowchart TD
    CAM["MaixCAM<br/>camera → NPU inference"] --> DL["Session log<br/>JSON on the SD card"]
    DL -->|"session ends"| UP["Uploader<br/>retries with backoff"]
    UP -->|"HTTPS · x-api-key<br/>POST /v1/events"| M["Companion Memory<br/>capture ingest"]

Setup

You need a MaixCAM or MaixCAM Pro, an SD card, and a Companion Memory instance the camera can reach over Wi-Fi if you want sync.

"network": {
  "upload_enabled": true,
  "endpoint_url": "https://companion.example.com",
  "api_key": "YOUR_CAPTURE_KEY"
}

Usage

• Start or stop a capture session — touch the top-left corner of the screen. While capturing, every enabled model runs on the live camera feed, and any detection is logged as a JSON event.
• Exit the app — touch the top-right corner.
• Status bar — shows the frame rate, the number of active models, the capture state, and, during a session, the events logged this session and in total.

When a session ends, its file is queued for upload if sync is on. Uploaded sessions move to an uploaded/ folder under the log path, compressed by default, and the app keeps the ten most recent session files.

What Companion Memory receives

Each detection is sent as one capture event with source_system set to ci-maixcam, a kind such as object_detection, a timestamp, a title, the detector’s own event as the payload, and the highest detection confidence when there is one. Companion Memory stores the payload opaquely, so you can find events by kind and title but not by the labels inside them. The camera never uploads images.

Troubleshooting

A model never produces events Check that its .mud file exists at the configured path. A missing file is logged and that model stays disabled while the app keeps running. If the file loads but nothing is detected, lower the model’s confidence_threshold in the configuration file.

Sync isn’t happening network.upload_enabled must be true and network.endpoint_url must be a real Companion Memory address. It is empty by default. Sync is skipped while Wi-Fi is disconnected and resumes on its own. Confirm the API key has the Capture scope: a rejected key logs an HTTP 401 or 403 and keeps retrying.

The app is slow, or runs out of memory Turn off models you don’t need on the setup screen. Only the models you enable are loaded.

Storage is filling up The app keeps only the ten most recent session files and archives uploaded sessions. Check the SD card’s free space if sessions keep piling up while sync is off.

Links

• Sipeed MaixCAM documentation 
• CCTV Camera Database  — open dataset of IP camera specifications; search, filter, and compare camera models
• Get your API keys
• Companion Memory
• How the Companion devices compare