Amodal Counting, Visibility-Corrected Counting (A4)
🔢
1
Visibility-corrected counting with calibrated intervals.
Dhi Labs demos
The six live Hugging Face Spaces demoing each Dhi Labs product end to end.
Note A4: compares a naive object count against the visibility corrected count and its calibrated interval on occluded scenes.
Multicam Reasoning Memory, Cross-Camera Identity (E1)
🧠Cross-camera linking with a precision-first refusal gate.
Note E1: links a simulated identity across two camera views using transit time priors, and shows when it refuses to link because it is unsure.
Fixed-Camera 3D, Metric 3D From One Camera (A3)
📐Metric 3D from one fixed camera, hard case disclosed.
Note A3: recovers metric 3D structure from a fixed camera by self calibrating against synthetic pedestrian trajectories, with no GPU or model weights involved.
Causal Predictive Alerting, Alerts That Grade Themselves (E4)
⏱Predictive alerting with a falsification ledger.
Note E4: replays a kinematic trajectory, raises a predictive alert before the incident, and shows the counterfactual reasoning plus the fulfilled or falsified verdict.
Thermal Perception, Honest About the Sensor Gap (A5)
🌡Thermal SSL harness, CPU-verified math, no fake benchmark.
Note A5: walks through the radiometric data engine on sample thermal imagery ahead of any GPU pretraining run.
Prompt2Model, A Vision Model Factory With a Refusal Gate (B1)
🏭A vision model factory that abstains and refuses to degrade.
Note B1: takes a natural language task description and walks through the pipeline stages toward a calibrated, quantized, exported model.