[Paper Review] A Strong and Reproducible Object Detector with Only Public Datasets

Tianhe Ren, Jianwei Yang+7 more|arXiv (Cornell University)|Apr 25, 2023

Advanced Neural Network Applications4 citations

TL;DR

This paper proposes Focal-Stable-DINO, a strong and fully reproducible object detector that achieves 64.6 AP on COCO val2017 and 64.8 AP on COCO test-dev using only public datasets and without test-time augmentation. It combines the FocalNet-Huge backbone with the Stable-DINO detector, leveraging only ImageNet-22K and Objects365 for pre-training, avoiding private data and complex training techniques.

ABSTRACT

This work presents Focal-Stable-DINO, a strong and reproducible object detection model which achieves 64.6 AP on COCO val2017 and 64.8 AP on COCO test-dev using only 700M parameters without any test time augmentation. It explores the combination of the powerful FocalNet-Huge backbone with the effective Stable-DINO detector. Different from existing SOTA models that utilize an extensive number of parameters and complex training techniques on large-scale private data or merged data, our model is exclusively trained on the publicly available dataset Objects365, which ensures the reproducibility of our approach.

Motivation & Objective

• To develop a strong yet fully reproducible object detection model that does not rely on private datasets or complex training techniques.
• To address the reproducibility crisis in object detection by using only publicly available datasets such as Objects365 and COCO.
• To improve detection performance on challenging object categories like small and crowded objects using a robust, publicly accessible pipeline.
• To highlight annotation quality issues in COCO, such as missing labels and inconsistent standards, which affect model generalization and training stability.

Proposed method

• Uses FocalNet-Huge (689M parameters) as the backbone, pre-trained only on ImageNet-22K, for strong feature representation.
• Employs the Stable-DINO detector, which introduces a position-supervised loss and position-modulated cost to stabilize training across decoder layers.
• Trains the full detector exclusively on the public Objects365 dataset for pre-training and COCO for fine-tuning, avoiding any private data.
• Avoids complex training techniques such as masked image modeling or image-text contrastive learning, ensuring simplicity and reproducibility.
• Uses a standard end-to-end training pipeline with standard data augmentation and standard optimization, without test-time augmentation.
• Relies on publicly available code and weights from GitHub repositories for FocalNet and Stable-DINO, enabling full reproducibility.

Experimental results

Research questions

• RQ1 Can a strong object detector be built using only publicly available datasets without relying on private data or complex pre-training techniques?
• RQ2 How does combining a powerful backbone (FocalNet-Huge) with a stable detection head (Stable-DINO) affect performance on COCO?
• RQ3 To what extent do annotation inconsistencies and missing labels in COCO impact model training and generalization?
• RQ4 Can a model achieve SOTA performance on COCO test-dev without test-time augmentation or large-scale data curation?
• RQ5 What are the performance disparities across object categories, especially for small and crowded objects, when using only public data?

Key findings

• Focal-Stable-DINO achieves 64.6 AP on COCO val2017 and 64.8 AP on COCO test-dev, setting a new SOTA for models trained exclusively on public datasets.
• The model outperforms existing SOTA methods that rely on private data, masked image modeling, or complex data curation pipelines.
• The model demonstrates robustness in detecting small and crowded objects, as shown in visualizations of hard cases with red arrows.
• The study identifies significant annotation issues in COCO, including missing labels for objects like apples and inconsistent labeling of items like bananas.
• The performance gap between object classes remains large, especially for small objects, indicating ongoing challenges in detection generalization.
• The model's success without test-time augmentation confirms its strong inductive bias and generalization capability, even under standard training protocols.