EduVision 2026: Classroom Crowd Detection & Student Counting

Background

Accurately detecting and counting people in real-world classroom environments is a key problem in smart campus automation, attendance monitoring, classroom utilization analytics, safety and crowd management. Unlike ideal datasets with single-person images, classroom scenes feature varying lighting conditions, occlusions, different viewing angles, diverse class sizes, and furniture shadows.

Core Objective

Participants must build a fully automated system that detects the location (bounding boxes) of all human subjects in classroom images and accurately estimates the number of students present.

• Classroom Crowd Detection
• Student Count Estimation

Dataset Description

The dataset includes classroom images from different angles, both small and large class sizes, varying lighting and background conditions, with manual annotations for people (bounding boxes and class labels).

Typical classroom scenes show anywhere from a handful to dozens of students, and may include chairs, desks, partial occlusions, and non-student persons.

Challenge Tasks

1. Person Detection

Develop models that detect all persons in each image, output bounding boxes, and distinguish between students and non-students (optional bonus).

Model outputs must follow standard object detection formats (e.g., COCO-style).

2. Student Counting

From detection outputs, compute a count of students present in the frame.

• Derived from detection model predictions
• Post-processing filters (size thresholds, context)
• Custom regression/counting algorithms

Evaluation Metrics

Detection Score (50%)

• Mean Average Precision (mAP)
• Based on IoU thresholds (0.5:0.95)

Counting Score (50%)

• Mean Absolute Error (MAE)
• MAE = average(abs(predicted_count - true_count))
• Lower is better

Rules & Constraints

Key Rules

• No external datasets unless explicitly allowed
• All models must run within provided test time and memory limits
• Pretrained weights are allowed if disclosed
• Use of public pre-trained backbones permitted
• Test set labels will be withheld for fair evaluation
• Participants must provide reproducible code

Deliverables

• Model weights and architecture details
• Inference code + instructions
• Predictions on the test dataset
• Detection bounding boxes
• Student count per image
• Technical report describing model design, data augmentation, post-processing logic, evaluation analysis