DeepCamera yolo-detection-2026

YOLO 2026 — state-of-the-art real-time object detection

install

source · Clone the upstream repo

git clone https://github.com/SharpAI/DeepCamera

Claude Code · Install into ~/.claude/skills/

T=$(mktemp -d) && git clone --depth=1 https://github.com/SharpAI/DeepCamera "$T" && mkdir -p ~/.claude/skills && cp -r "$T/skills/detection/yolo-detection-2026" ~/.claude/skills/sharpai-deepcamera-yolo-detection-2026 && rm -rf "$T"

manifest: skills/detection/yolo-detection-2026/SKILL.md

source content

YOLO 2026 Object Detection

Real-time object detection using the latest YOLO 2026 models. Detects 80+ COCO object classes including people, vehicles, animals, and everyday objects. Outputs bounding boxes with labels and confidence scores.

Model Sizes

Size	Speed	Accuracy	Best For
nano	Fastest	Good	Real-time on CPU, edge devices
small	Fast	Better	Balanced speed/accuracy
medium	Moderate	High	Accuracy-focused deployments
large	Slower	Highest	Maximum detection quality

Hardware Acceleration

The skill uses

env_config.py

to automatically detect hardware and convert the model to the fastest format for your platform. Conversion happens once during deployment and is cached.

Platform	Backend	Optimized Format	Compute Units	Expected Speedup
NVIDIA GPU	CUDA	TensorRT `.engine`	GPU	~3-5x
Apple Silicon (M1+)	MPS	CoreML `.mlpackage`	Neural Engine (NPU)	~2x
Intel CPU/GPU/NPU	OpenVINO	OpenVINO IR `.xml`	CPU/GPU/NPU	~2-3x
AMD GPU	ROCm	ONNX Runtime	GPU	~1.5-2x
CPU (any)	CPU	ONNX Runtime	CPU	~1.5x

Apple Silicon Note: Detection defaults to
cpu_and_ne
(CPU + Neural Engine), keeping the GPU free for LLM/VLM inference. Set
compute_units: all
to include GPU if not running local LLM.

How It Works

deploy.sh

detects your hardware via

env_config.HardwareEnv.detect()

Installs the matching
```
requirements_{backend}.txt
```
(e.g. CUDA → includes
```
tensorrt
```
)
Pre-converts the default model to the optimal format
At runtime,
```
detect.py
```
loads the cached optimized model automatically
Falls back to PyTorch if optimization fails

Set

use_optimized: false

to disable auto-conversion and use raw PyTorch.

Auto Start

Set

auto_start: true

in the skill config to start detection automatically when Aegis launches. The skill will begin processing frames from the selected camera immediately.

auto_start: true
model_size: nano
fps: 5

Performance Monitoring

The skill emits

perf_stats

events every 50 frames with aggregate timing:

{"event": "perf_stats", "total_frames": 50, "timings_ms": {
  "inference": {"avg": 3.4, "p50": 3.2, "p95": 5.1},
  "postprocess": {"avg": 0.15, "p50": 0.12, "p95": 0.31},
  "total": {"avg": 3.6, "p50": 3.4, "p95": 5.5}
}}

Protocol

Communicates via JSON lines over stdin/stdout.

Aegis → Skill (stdin)

{"event": "frame", "frame_id": 42, "camera_id": "front_door", "timestamp": "...", "frame_path": "/tmp/aegis_detection/frame_front_door.jpg", "width": 1920, "height": 1080}

Skill → Aegis (stdout)

{"event": "ready", "model": "yolo2026n", "device": "mps", "backend": "mps", "format": "coreml", "gpu": "Apple M3", "classes": 80, "fps": 5}
{"event": "detections", "frame_id": 42, "camera_id": "front_door", "timestamp": "...", "objects": [
  {"class": "person", "confidence": 0.92, "bbox": [100, 50, 300, 400]}
]}
{"event": "perf_stats", "total_frames": 50, "timings_ms": {"inference": {"avg": 3.4}}}
{"event": "error", "message": "...", "retriable": true}

Bounding Box Format

[x_min, y_min, x_max, y_max]

— pixel coordinates (xyxy).

Stop Command

{"command": "stop"}

Installation

The

deploy.sh

bootstrapper handles everything — Python environment, GPU backend detection, dependency installation, and model optimization. No manual setup required.

./deploy.sh

Requirements Files

File	Backend	Key Deps
`requirements_cuda.txt`	NVIDIA	`torch` (cu124), `tensorrt`
`requirements_mps.txt`	Apple	`torch` , `coremltools`
`requirements_intel.txt`	Intel	`torch` , `openvino`
`requirements_rocm.txt`	AMD	`torch` (rocm6.2), `onnxruntime-rocm`
`requirements_cpu.txt`	CPU	`torch` (cpu), `onnxruntime`