Digit Detection Problem - RT-DETRv2 🚀

NYCU Visual Recognition using Deep Learning (Spring 2026) - Homework 2

This repository contains the official implementation for the Digit Detection Problem (Homework 2). By leveraging the RT-DETRv2 architecture with a ResNet-50 backbone, this project successfully secured high rank on the NYCU Visual Recognition Course (Spring 2026) Homework 2 CodaBench competition leaderboard with a Test score of 0.41.

---

📌 Project Summary

The core objective of this project is to localize and classify multiple digits in dense environments. To maximize performance under strict course constraints, we implemented several key strategies:

• Rectangular Training ($320 \times 640$): We adapted the input resolution to preserve the widescreen aspect ratio of the dataset, preventing vertical feature distortion.
• Bipartite Matching Loss Analysis: We explored the impact of the End-of-Sequence (eos_coef) background penalty to understand the balance between local precision and test-set generalization.
• Advanced Regularization: Implementation of Random Erasing, OneCycleLR scheduling, and Exponential Moving Average (EMA) for model weights to ensure stable convergence.

---

📊 Results & Experimental Analysis

We evaluated 14 distinct configurations to bridge the gap between local validation and the unseen CodaBench test set.

Key Finding: While increasing the background penalty (eos_coef = 1e-3) yielded a superior local mAP of 0.47, it caused a performance drop on CodaBench. This suggests that a more balanced regularization approach is necessary to handle noisy annotations in the test set.

---

📈 Training Dynamics

The learning curves below demonstrate the training progression of our optimized configuration.

---

🛠️ Usage Guide

1. Requirements

Install the necessary packages via pip:

pip install -r requirements.txt

2. How to Run Training

There are two primary methods to execute training depending on your hardware environment.

Option A: Using train.py (Direct Execution)

Best for single-GPU setups or debugging. This provides direct access to Python's error logs and single-process execution.

python train.py --run_name --batch_size --epochs --lr --wd --queries --gpu --workers --eof --data_path --save_path

Option B: Using train.sh (Distributed Execution)

Best for Multi PC/Container setups. Generate output name based on the parameters it used.

bash train.sh --batch_size --epochs --lr --wd --queries --gpu --workers --eof

🔍 Differences: train.py vs train.sh

---

👤 Author

Muhammad Rayhan Athaillah (賴瑞涵)
Student ID: 313540001
Affiliation: National Yang Ming Chiao Tung University (NYCU)