Visual Perception â€" MITCF Evaluation
Qwen2.5-VL-3B-Instruct evaluated on POPE + MME benchmarks under four paired conditions. Targeting Elsevier Displays Special Issue: "Perceptual AI Systems in the Era of Large Models."
Experimental Conditions
Baseline
Unmodified Qwen2.5-VL-3B, no intervention
Revision Only
MITCF verification + critique + revision, no fuzzy gate
Fixed Gate
Fixed-threshold confidence gate comparator
Full MITCF
Complete Mamdani fuzzy-gated MITCF pipeline
POPE Evaluation
Object hallucination â€" random, popular, adversarial
MME Evaluation
14-category visual perception with official scoring
Condition Comparison
Side-by-side metrics for all four conditions
Trustworthiness Index
TI: hallucination + bias + robustness + decision
Adversarial Robustness
Perturbation degradation curves
Run VP Evaluation
Launch evaluate / robustness / ablation jobs
Dataset Setup
POPE + MME download status and instructions
Recent VP Runs
+ Launch new runLoading…