Research map / Image generation / Autoregressive and latent generation
DISCount: Counting in Large Image Collections with Detector-Based Importance Sampling
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DISCount, a detector-based importance sampling framework with human screening, multi-region k-DISCount estimators, control variates, and confidence intervals to count objects in large image collections with fewer labels.
Problem
Counting objects in massive fixed image collections when an imperfect detector cannot give accurate counts, so human screening effort must be allocated to produce unbiased, low-error count estimates for multiple overlapping regions.
Approach
Use detector outputs as an importance-sampling proposal over the whole domain, sample once and assign samples to many regions (post-stratification style), estimate unbiased (conditionally) counts by averaging human-verified/detector weight ratios, add control variates from calibrated detector or prior-year counts, and build asymptotic confidence intervals.
Key results
- DISCount reaches 20% error with about 1.6x fewer samples than IS-Count and about 3x fewer than simple Monte Carlo on both tasks.
- In labeling-effort terms, DISCount needs 6x less effort than IS-Count and 9x less than MC for bird counting, and 8x less than IS-Count and 12x less than MC for building counting.
- Confidence interval coverage stays below the nominal 95% for all methods; k-DISCount-cv-sigma(Omega) reaches up to about 80% coverage.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use detector-based importance sampling instead of covariate-based IS-Count | Count estimation on bird and building tasks · Bird and building counting tasks | samples needed to reach 20% error: ≈ 1.6× fewer samples | Helped |
| Use detector-based importance sampling instead of simple Monte Carlo | Count estimation on bird and building tasks · Bird and building counting tasks | samples needed to reach 20% error: ≈ 3× fewer samples | Helped |
| Verify detector outputs instead of annotating images from scratch | DISCount · Roosting birds | labeling effort at 20% error: 6× less effort than IS-Count; 9× less effort than MC | Helped |
| Verify detector outputs instead of annotating images from scratch | DISCount · Palu Tsunami damaged buildings | labeling effort at 20% error: 8× less effort than IS-Count; 12× less effort than MC | Helped |
| Estimate multiple cumulative regions jointly rather than running DISCount separately (ablation) | Bird counting · Cumulative quarterly counts | estimation error | Helped |
| Estimate multiple quarterly regions jointly rather than running DISCount separately (ablation) | Bird counting · Individual quarterly counts | estimation error | Mixed |
| Estimate multiple cumulative regions jointly rather than running DISCount separately (ablation) | Building damage counting · Palu Tsunami cumulative regions | estimation error | Helped |
| Add calibrated detector counts as control variates (ablation) | k-DISCount · Roosting birds | estimation error | Helped |
| Add calibrated detector counts as control variates (ablation) | k-DISCount · Cumulative daily bird counts | confidence interval width | Helped |
| Estimate variance using all samples rather than only samples in each region (ablation) | k-DISCount-cv · Cumulative daily bird counts | confidence interval coverage: up to ≈ 80% coverage | Helped |
| Use isotonic calibration instead of DISCount | Bird-count estimation · Roosting birds | estimation error | Hurt |
Surprising findings
- Unscreened detector counts give roughly 136% and 149% error for building and bird counting, showing the detector alone is unreliable.
- Confidence interval coverage is substantially below nominal for all methods, including the proposed one.
Limitations
- Approach depends on availability of a reasonably good detector.
- Confidence interval coverage is slightly low.
- Labeling-effort savings rely on an assumed verification-to-annotation time ratio taken from ImageNet studies.
Experimental setup
Datasets: xBD NEXRAD weather radar (Great Lakes screened detections)
Benchmarks: xBD Palu Tsunami region NEXRAD Great Lakes bird roost counts
Metrics: fractional count error confidence interval width confidence interval coverage labeling effort
Baselines: Simple Monte Carlo IS-Count Isotonic regression calibration Naive detector without screening
Builds on: IS-Count Neyman allocation active testing
Closest papers by meaning
- See What We Cannot See: A Geo-guided Reasoning Benchmark for Object Counting under Adverse Earth Observation Conditions (CVPR 2026 · similarity 0.85)
- DAVE - A Detect-and-Verify Paradigm for Low-Shot Counting (CVPR 2024 · similarity 0.84)
- Generalizable Disaster Damage Assessment via Change Detection with Vision Foundation Model (AAAI 2025 · similarity 0.83)
- Benchmarking Object Detectors under Real-World Distribution Shifts in Satellite Imagery (CVPR 2025 · similarity 0.83)
- InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition (NeurIPS 2025 · similarity 0.83)
- COUNTS: Benchmarking Object Detectors and Multimodal Large Language Models under Distribution Shifts (CVPR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.