ON-DEMAND WEBINAR · 40 MIN
Jump to a section
Pick a chapter and the video plays from that point.
Wondering how this applies to your own pipeline?
≋
LANE DETECTION
97%
less source storage at CQ 35, while CLRNet kept 97.4% of its original lane predictions.
□
2D OBJECT DETECTION
0.96 mAP
RF-DETR detections on CABR-optimized video vs. the original, across ~200,000 objects (mean IoU 0.97).
≡
VLM CAPTIONING
41–57%
lower bitrate, with caption changes smaller than the model’s own run-to-run variation. A proof of concept with Qwen2.5-VL.
Evaluate your data
SPEAKERS
Tamar Shoham
CTO
Ronen Nissim
AI Team Lead
David Tomarov
Computer Vision Researcher
Ethan Fenakel
EVALUATE COMPRESSION ON YOUR DATA
Tell us about your models and video setup. We’ll show you where compression starts to change their output.
✓
How much your current pipeline can compress safely
✓
Which of your models are most sensitive
✓
Where ML-Safe Compression fits — or doesn’t