Release 2025/08/27
v3.6.2
Observ Aug Updates
We’re excited to ship a set of upgrades focused on smarter model usage, video analysis, and cutting compute on static scenes.
New Features:
✅ Task-Specific Model Selection
Assign a specific model (and version) per task. This lets you tune accuracy vs. cost—for example, lightweight models for low-resolution feeds and stronger models for high-precision scenarios—without affecting other tasks.
✅ Video Detection
A brand-new, frame-by-frame inference pipeline that turns uploaded videos into timestamped detections and reviewable result clips. Use it to apply consistent settings across videos, enrich events with semantics (via VLM), and audit outcomes in one place.
Batch uploads within a single task: Upload a set of different videos to the same video task and run them with the same model and parameters for consistent, apples-to-apples detection.
VLM-powered analysis: Apply a Vision-Language Model to improve semantic understanding of events in video.
Result playback: Watch rendered videos with detection overlays immediately after processing.
Analysis & History integration: Review results and related statistics in the Analysis page and track past runs in History.
✅ Computer Vision Filters (image quality gating)
Filter out low-quality or ambiguous events before they reach reviewers or downstream automations. Supported criteria include Resolution, FFT-Blur, and BRISQUE, helping reduce false alarms from blurry or unclear footage.
✅ Caching Inference (similar-frame skipping)
Detect redundant frames by visual similarity and skip repeated inference to significantly reduce compute—ideal for fixed, rarely changing camera views.
🚀 Streaming Performance Optimizations
More under-the-hood improvements have been made to enhance real-time streaming performance, improving decoding speed, memory usage, and overall system responsiveness—especially under high-load or large scale camera deployments.
Coming Soon:
We’re preparing to roll out powerful new features to give you even more flexibility and control over task execution and resource usage:
VLM Flow (Stepwise & Token Optimization) Soon, you’ll be able to route VLM tasks through a guided, multi-stage flow. This approach reduces unnecessary context expansion, cutting overall token usage while improving reliability and traceability.
Multi-Image VLM Input You’ll be able to submit multiple images (e.g., keyframes or an image set) to a single VLM query. This enables cross-frame consistency checks, better object re-identification, and richer scene understanding in one pass.
INT8-Optimized Models We’re adding INT8-quantized model variants (TensorRT). Expect higher throughput and lower memory usage—ideal for high-volume deployments—while maintaining competitive accuracy.
Stay tuned—these upgrades are just around the corner and will further enhance the scalability and efficiency of Observ in real-world deployments.
Support
For technical support, please contact our support team during business hours.
Last updated