[Paper Review] A framework for robust object multi-detection with a vote aggregation and a cascade filtering

Grzegorz Kurzejamski, J. Zawistowski+1 more|arXiv (Cornell University)|Dec 29, 2015

Advanced Image and Video Retrieval Techniques18 references3 citations

TL;DR

This paper presents a robust, parameter-optimized framework for multi-object detection in retail shelf scenes using SIFT features, vote aggregation, and a cascade of filters. It achieves 89% detection rate with under 1% false detection chance using only one pattern per object and minimal manual tuning, outperforming prior methods in low-resolution and complex lighting conditions.

ABSTRACT

This paper presents a framework designed for the multi-object detection purposes and adjusted for the application of product search on the market shelves. The framework uses a single feedback loop and a pattern resizing mechanism to demonstrate the top effectiveness of the state-of-the-art local features. A high detection rate with a low false detection chance can be achieved with use of only one pattern per object and no manual parameters adjustments. The method incorporates well known local features and a basic matching process to create a reliable voting space. Further steps comprise of metric transformations, graphical vote space representation, two-phase vote aggregation process and a cascade of verifying filters.

Motivation & Objective

• To develop a scalable, robust multi-object detection system for retail product search on shelves with minimal manual parameter tuning.
• To address challenges in detecting generic graphics with high visual similarity across brands and product variations.
• To achieve high detection rates and low false positive rates despite complex lighting, reflections, and partial occlusions.
• To enable effective detection using only one pattern per object, avoiding costly data collection and learning processes.
• To improve reliability through dynamic parameterization and a two-phase filtering cascade.

Proposed method

• The system uses SIFT features for local feature extraction and matching between scene and pattern images.
• A vote space is constructed by aggregating geometrically consistent matches using RANSAC-like transformations.
• Graphical vote space representation enables visualization and filtering of high-confidence object locations.
• A two-phase vote aggregation process enhances localization accuracy by refining candidate detections.
• A cascade of verifying filters rejects false positives by applying metric thresholds and pattern-specific validation.
• Dynamic parameterization adjusts detection sensitivity based on scene complexity and pattern quality.

Experimental results

Research questions

• RQ1 How can a multi-detection system achieve high detection rates with minimal false positives using only one pattern per object?
• RQ2 What role does dynamic parameterization play in improving detection robustness across varying image qualities and lighting?
• RQ3 How does vote aggregation and cascaded filtering compare to traditional sliding window or KNN-based approaches in complex retail scenes?
• RQ4 To what extent does pattern quality (e.g., sharpness, noise, lighting) affect detection performance in a one-pattern-per-object setup?
• RQ5 Can a non-learning, feature-based approach outperform learning-based methods in detecting generic graphics with high visual redundancy?

Key findings

• The system achieved a detection rate of 89.0% on 12 MPx images with a false detection chance of 0.72% per image.
• At 3 MPx resolution, the detection rate was 84.4%, with a slightly higher false detection chance of 1.63%, indicating resolution-dependent performance trade-offs.
• Despite using a 2-phase processing approach that multiplies false detections, the system maintained a low false positive rate due to effective filtering.
• The average number of false detections per image was 3.07 at 12 MPx and 3.28 at 3 MPx, showing that false positives are manageable with filtering.
• Pattern quality had a dominant impact on performance, with poor-quality patterns reducing detection rates to 0% even with optimal parameters.
• The framework demonstrated linear scalability with respect to the number of patterns and was effective even in scenes with strong reflections and partial occlusions.