Source-linked AI summary

DORAEMON: A Unified Library for Visual Object Modeling and Representation Learning at Scale

Ke Du, Yimin Peng, Chao Gao, Fan Zhou, Siqiao Xue

arXiv:2511.04394v1cs.CV

TL;DR

Visual modeling research is hindered by fragmented codebases, task-specific pipelines, and inconsistent training practices. DORAEMON addresses this gap with a unified YAML-driven PyTorch framework built around shared backbones, modular heads, configurable training, and deployment integrations. It provides broad, ready-to-use support for classification, face recognition, and retrieval while remaining extensible for large-scale workloads.

  • Problem

    Fragmented codebases, task-specific pipelines, and inconsistent training practices make replication and cross-task comparison difficult.

  • Method

    DORAEMON unifies shared backbones, modular task heads, configurable augmentation and losses, pretrained models, and training utilities in a YAML-driven PyTorch framework.

  • Results

    DORAEMON delivers strong, ready-to-use baselines for image classification, face recognition, and content-based image retrieval.

  • Takeaways & Limitations

    The library lowers the entry barrier for computer-vision research while remaining extensible for large-scale production workloads.

Abstract

from arXiv · show

DORAEMON is an open-source PyTorch library that unifies visual object modeling and representation learning across diverse scales. A single YAML-driven workflow covers classification, retrieval and metric learning; more than 1000 pretrained backbones are exposed through a timm-compatible interface, together with modular losses, augmentations and distributed-training utilities. Reproducible recipes match or exceed reference results on ImageNet-1K, MS-Celeb-1M and Stanford online products, while one-command export to ONNX or HuggingFace bridges research and deployment. By consolidating datasets, models, and training techniques into one platform, DORAEMON offers a scalable foundation for rapid experimentation in visual recognition and representation learning, enabling efficient transfer of research advances to real-world applications. The repository is available at https://github.com/wuji3/DORAEMON.

1 Introduction

Large-scale visual object modeling is growing, but fragmented codebases and inconsistent task-specific practices make replication and cross-task comparison difficult. DORAEMON addresses this need with a unified PyTorch framework offering broad model support, configurable training, and deployment-oriented tooling.

  • 1 Introduction: Fragmented codebases, task-specific pipelines, and inconsistent training practices complicate replication, integration, and cross-task comparison.Researchers must manually combine datasets, architectures, and loss functions, slowing experimentation.
  • 1 Introduction: DORAEMON provides a unified PyTorch framework for diverse large-scale visual object-modeling tasks.
  • 1 Introduction: Over 1,000 pretrained architectures and configurable optimization, regularization, augmentation, learning-rate, and hard-example-mining components support broad benchmarking and fine-grained training control.
  • 1 Introduction: Built-in Grad-CAM supports qualitative representation analysis and debugging, while HuggingFace integration provides a deployment interface.
  • 1 Introduction: Figure 1 tracks cumulative arXiv publications and GitHub projects in large-scale visual object modeling from 2019 to 2024.

2 Related Work

Existing libraries improved access to backbones or expanded task coverage, but important gaps remained in unified scope, maintenance, and end-to-end research-to-deployment support. DORAEMON combines multi-task workflows, extensive pretrained backbones, distributed training, and export capabilities in one YAML-driven framework.

  • 2 Related Work: Earlier libraries standardized backbones or supported individual task families, but researchers still faced bespoke pipelines, limited scope, or slowed maintenance.
  • 2 Related Work: DORAEMON unifies image classification, retrieval, and metric learning through YAML-driven workflows with more than 1,000 pretrained backbones.
  • 2 Related Work: Elastic fault-tolerant distributed training and one-click export to TensorRT, ONNX, or TorchScript connect research workflows with production deployment.

3 Technical Background

Visual representation learning extracts reusable embeddings from images and adapts them to different tasks through lightweight task-specific heads. DORAEMON applies this shared-backbone, modular-head paradigm across classification, face recognition, and image retrieval.

  • 3 Technical Background: A parameterized encoder maps an image x to a compact, discriminative, reusable embedding z = fθ(x) ∈ R^d for downstream tasks.
  • 3 Technical Background: Decoupling the shared encoder from lightweight task heads enables unified training, joint optimization, and scalable task adaptation.
  • 3 Technical Background: Classification, face recognition, and retrieval use different objectives but share the need for rich, transferable visual representations.
  • 3 Technical Background: DORAEMON integrates more than 1,000 pretrained models through timm and supports consistent pretraining, fine-tuning, and evaluation across the three task families.
  • 3 Technical Background: Its unified backbone with flexible heads is intended to streamline model development and support joint optimization and cross-task generalization.

4 Pipeline

DORAEMON implements a modular training pipeline that processes data, extracts representations with a shared backbone, adapts them through task-specific heads, and computes configurable losses. YAML controls augmentation, optimization, regularization, evaluation, and scalable integration across tasks and datasets.

  • 4 Pipeline: The pipeline combines data processing, shared representation extraction, task adaptation, and loss computation for classification, face recognition, and retrieval.
  • 4 Pipeline: Task-agnostic preprocessing supports resizing, dynamic augmentations, label handling, and curriculum-style augmentation schedules controlled by YAML flags.
  • 4 Pipeline: A timm-based shared encoder supports ResNet, Swin Transformer, Vision Transformer, MAE, and CLIP variants totaling over 1,000 pretrained models.
  • 4 Pipeline: Modular heads pair classification with softmax and cross-entropy, face recognition with angular-margin objectives, and retrieval with triplet-based or contrastive objectives.
  • 4 Pipeline: Optimizers, schedulers, regularizers, and task-specific metrics such as Top-1 accuracy, Recall@K, and ROC AUC are configurable and logged automatically.
  • 4 Pipeline: The system supports scalable multi-task or multi-dataset sessions, PyTorch checkpoint export, and HuggingFace-compatible APIs.

5 Software Interface

DORAEMON provides a unified and extensible framework for large-scale visual object modeling and representation learning across classification, retrieval, and metric learning. It supports modular integration and benchmarking of backbones, losses, and training strategies for research and industrial use.

  • DORAEMON unifies visual object modeling and representation learning across classification, retrieval, and metric learning.
  • Researchers can integrate and benchmark new backbones, losses, and training strategies with minimal effort through its modular interface.
  • The framework is designed to facilitate fair comparisons with existing methods across visual-learning tasks.
  • DORAEMON also provides configuration and tooling intended for industrial practitioners.

6 Ongoing and Future Work

DORAEMON’s future work extends the library toward tool-using vision agents and multimodal large-language-model training. Planned capabilities include database integration, temporal vision analytics, continual multimodal pre-training, and prompt-based tuning.

  • The roadmap proposes integrating powerful agents that can connect to databases and support time-stamped vision analytics over historical image and video streams.
  • These additions would extend DORAEMON from a training toolkit toward a tool-using agent for vision.
  • Planned multimodal training support includes continual pre-training that incrementally ingests new text–image or video data.
  • The roadmap also includes lightweight prompt-based tuning across modalities for prototyping lifelong and task-specific multimodal LLMs.

7 Conclusion

DORAEMON delivers ready-to-use baselines and extensible tooling for several visual-learning tasks. Its roadmap positions the library as a springboard for multimodal large-language-model research and broader community-driven development.

  • DORAEMON delivers strong, ready-to-use baselines for image classification, face recognition, and content-based image retrieval.
  • Its modular APIs, data-augmentation suite, and visualization tools lower the entry barrier for computer-vision research.
  • The library remains extensible for large-scale production workloads.
  • The roadmap positions DORAEMON as a springboard for multimodal large-language-model research and encourages community adoption and extension.
Loading 2511.04394v1…