@researai/deepscientist 1.5.16 → 1.6.0

package/AISB/catalog/aisb.t3.056_infosam.yaml

@@ -0,0 +1,224 @@
+id: aisb.t3.056_infosam
+name: 'InfoSAM: Fine-Tuning the Segment Anything Model from An Information-Theoretic
+  Perspective'
+version: 0.1.0
+one_line: Information-theoretic SAM fine-tuning via mutual-information-based distillation
+  preserving domain-invariant relational knowledge for specialized segmentation.
+task_description: 'InfoSAM is a parameter-efficient fine-tuning (PEFT) framework for
+  SAM that enhances segmentation performance in specialized domains by preserving
+  pre-trained domain-invariant knowledge through information-theoretic distillation.
+  The method introduces two novel mutual information-based objectives: (1) a relational
+  compression loss that extracts domain-invariant structural patterns from pre-trained
+  SAM by minimizing mutual information between task-irrelevant features and the relational
+  module outputs, and (2) a cross-model distillation loss that maximizes mutual information
+  between teacher (pre-trained SAM) and student (fine-tuned SAM) relational representations.
+  The approach uses attention-driven relation modules and Renyi entropy-based quantification
+  with normalized Gram matrices computed via Frobenius norm. Benchmarked across diverse
+  domains including natural images, medical imaging, agriculture (leaf disease), and
+  remote sensing.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: external_eval_required
+support_level: recovery
+time_band: 1-2d
+cost_band: high
+difficulty: hard
+data_access: public
+primary_outputs:
+- s_measure
+- e_measure
+- weighted_f_measure
+- infosam_checkpoint
+- segmentation_results
+launch_profiles:
+- id: train_only
+  label: Train Only
+  description: Run InfoSAM fine-tuning with information-theoretic distillation loss
+    on specified dataset.
+- id: eval_only
+  label: Evaluation Only
+  description: Evaluate fine-tuned model using S-measure, E-measure, and weighted
+    F-measure.
+- id: full_pipeline
+  label: Full Pipeline
+  description: Complete training and evaluation pipeline with checkpoint management.
+dataset_download:
+  primary_method: mixed
+  sources:
+  - kind: github_repo
+    url: https://github.com/autogluon/autogluon/tree/master/examples/automm/Conv-LoRA
+    access: public
+    note: Datasets sourced from Conv-LoRA repository covering multiple segmentation
+      domains
+  - kind: huggingface_model
+    url: https://huggingface.co/MuyaoYuan/InfoSAM/tree/main
+    access: public
+    note: Pre-trained InfoSAM checkpoints available for direct use
+  - kind: facebook_ai
+    url: https://dl.fbaipublicfiles.com/segment_anything/
+    access: public
+    note: Pre-trained SAM teacher models (ViT-B, ViT-H, ViT-L) required for training
+  notes:
+  - The current local snapshot does not include the `data/data_download.py` script
+    referenced in README
+  - Dataset acquisition requires reconstruction from Conv-LoRA external route
+  - Shell scripts `config/train/InfoSAM_b_leaf.sh` and `config/eval/InfoSAM_b_leaf.sh`
+    referenced in README are absent from snapshot
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - No explicit API credentials or dataset access tokens detected in bundled materials
+  - HuggingFace model access is public
+  - Facebook AI SAM checkpoint downloads are public
+resources:
+  minimum:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 24
+  recommended:
+    cpu_cores: 32
+    ram_gb: 128
+    disk_gb: 200
+    gpu_count: 1
+    gpu_vram_gb: 48
+environment:
+  python: '3.8'
+  cuda: '12.1'
+  pytorch: 2.4.1
+  flash_attn: null
+  key_packages:
+  - torchvision==0.19.1
+  - transformers
+  - pillow
+  - numpy
+  - opencv-python
+  - scipy
+  notes:
+  - Tested environment: Python 3.8.20, PyTorch 2.4.1, torchvision 0.19.1, CUDA 12.1
+      on Ubuntu 20.04
+  - Environment specifications derived from README documentation; end-to-end execution
+    requires restoration of missing launcher scripts
+risk_flags:
+- source_snapshot_incomplete
+- launch_scripts_missing
+- dataset_download_script_absent
+risk_notes:
+- Current snapshot preserves documentation (README, OPTIMIZATION.md, AGENTS.md, latex.md)
+  but lacks executable train/eval launcher trees
+- Missing `data/data_download.py` script required for dataset preparation
+- Missing `config/train/InfoSAM_b_leaf.sh` and `config/eval/InfoSAM_b_leaf.sh` for
+  training and evaluation
+- Treat as paper-backed reconstruction target; full execution requires upstream script
+  restoration from official repository
+recommended_when: 'Use this benchmark to reconstruct and audit the InfoSAM PEFT methodology
+  for SAM adaptation. Suitable for researchers investigating information-theoretic
+  distillation for vision foundation models, relational knowledge transfer, and domain-invariant
+  segmentation. Prepare to restore missing training and evaluation scripts from the
+  official repository.
+
+  '
+not_recommended_when: 'Do not use as a ready-to-run local benchmark. The current snapshot
+  is incomplete relative to the official README. Full execution requires restoring
+  missing launcher scripts, download utilities, and configuration files from the upstream
+  InfoSAM repository before the benchmark becomes operational.
+
+  '
+paper:
+  title: 'InfoSAM: Fine-Tuning the Segment Anything Model from An Information-Theoretic
+    Perspective'
+  venue: ICML 2025 Spotlight
+  year: 2025
+  url: https://proceedings.mlr.press/v267/zhang25ct.html
+  arxiv_url: https://arxiv.org/pdf/2505.21920
+  project_page: https://muyaoyuan.github.io/InfoSAM_Page
+  authors:
+  - Yuanhong Zhang
+  - Muyao Yuan
+  - Weizhan Zhang
+  - Tieliang Gong
+  - Wen Wen
+  - Jiangyong Ying
+  - Weijie Shi
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.056_infosam.zip
+  archive_type: zip
+  local_dir_name: paper-56-InfoSAM
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.056_infosam.zip
+  sha256: 6339252a6968df0ec70cdac20e1eb44e1eeae9f33b3b6386513a2772d1e51ba9
+  size_bytes: 45688
+display:
+  palette_seed: gold-forest-segment
+  art_style: info-geometry
+  accent_priority: high
+  domain_tags:
+  - computer_vision
+  - segmentation
+  - foundation_models
+  - distillation
+  - parameter_efficient_finetuning
+  - information_theory
+image_path: ../../../AISB/image/056_aisb.t3.056_infosam.jpg
+capability_tags:
+- research_code_optimization
+- computer_vision
+- segmentation
+- foundation_models
+- distillation
+- parameter_efficient_finetuning
+- mutual_information
+- knowledge_distillation
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+metric_contracts:
+  s_measure:
+    description: Structure measure for foreground segmentation quality evaluation
+    origin: paper_evaluation
+    blocked_reason: No executable evaluation code anchor in current snapshot
+  e_measure:
+    description: Enhanced alignment measure for boundary and region evaluation
+    origin: paper_evaluation
+    blocked_reason: No executable evaluation code anchor in current snapshot
+  weighted_f_measure:
+    description: Weighted F-measure combining precision and recall with optimal threshold
+      selection
+    origin: paper_evaluation
+    blocked_reason: No executable evaluation code anchor in current snapshot
+external_models:
+  teacher_models:
+  - model_id: sam_vit_b
+    url: https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth
+    description: ViT-Base SAM model (default teacher)
+  - model_id: sam_vit_h
+    url: https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth
+    description: ViT-Huge SAM model
+  - model_id: sam_vit_l
+    url: https://dl.fbaipublicfiles.com/segment_anything/sam_vit_l_0b3195.pth
+    description: ViT-Large SAM model
+  pretrained_checkpoints:
+  - model_id: infosam_finetuned
+    url: https://huggingface.co/MuyaoYuan/InfoSAM/tree/main
+    description: Pre-trained InfoSAM checkpoints on HuggingFace
+    access: public
+restoration_required:
+- data/data_download.py
+- config/train/InfoSAM_b_leaf.sh
+- config/eval/InfoSAM_b_leaf.sh
+- requirements.txt
+- launcher scripts referenced in README
+restoration_notes: 'The official repository contains complete training and evaluation
+  scripts. Reconstruction involves cloning the upstream repository or restoring the
+  referenced shell scripts and download utilities to achieve a fully operational benchmark.
+
+  '

package/AISB/catalog/aisb.t3.056_infosam.zh.yaml

@@ -0,0 +1,189 @@
+id: aisb.t3.056_infosam
+name: 'InfoSAM: Fine-Tuning the Segment Anything Model from An Information-Theoretic
+  Perspective'
+version: 0.1.0
+one_line: |
+  基于信息论的SAM微调方法，通过互信息蒸馏保留领域无关的关系知识，用于专业分割任务。
+task_description: |
+  InfoSAM是一种用于SAM的参数高效微调（PEFT）框架，通过信息论蒸馏保留预训练领域无关知识，从而提升专业领域的分割性能。该方法引入两个基于互信息的新目标：（1）关系压缩损失，通过最小化任务无关特征与关系模块输出之间的互信息，从预训练SAM中提取领域无关的结构模式；（2）跨模型蒸馏损失，最大化教师（预训练SAM）与学生（微调后SAM）关系表示之间的互信息。该方法使用注意力驱动的关系模块和基于Renyi熵的量化，通过Frobenius范数计算的归一化Gram矩阵。在包括自然图像、医学影像、农业（叶片病害）和遥感在内的多个领域进行了基准测试。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: external_eval_required
+support_level: recovery
+time_band: 1-2d
+cost_band: high
+difficulty: hard
+data_access: public
+primary_outputs:
+- s_measure
+- e_measure
+- weighted_f_measure
+- infosam_checkpoint
+- segmentation_results
+launch_profiles:
+- id: train_only
+  label: 仅训练
+  description: 使用信息论蒸馏损失在指定数据集上运行InfoSAM微调。
+- id: eval_only
+  label: 仅评估
+  description: 使用S-measure、E-measure和加权F-measure评估微调后的模型。
+- id: full_pipeline
+  label: 完整流程
+  description: 包含检查点管理的完整训练和评估流程。
+dataset_download:
+  primary_method: mixed
+  sources:
+  - kind: github_repo
+    url: https://github.com/autogluon/autogluon/tree/master/examples/automm/Conv-LoRA
+    access: public
+    note: 数据集来源于Conv-LoRA仓库，覆盖多个分割领域
+  - kind: huggingface_model
+    url: https://huggingface.co/MuyaoYuan/InfoSAM/tree/main
+    access: public
+    note: 预训练的InfoSAM检查点，可直接使用
+  - kind: facebook_ai
+    url: https://dl.fbaipublicfiles.com/segment_anything/
+    access: public
+    note: 训练所需的预训练SAM教师模型（ViT-B、ViT-H、ViT-L）
+  notes:
+  - 当前本地快照不包含README中引用的`data/data_download.py`脚本
+  - 数据集获取需要从Conv-LoRA外部路径重建
+  - README中引用的shell脚本`config/train/InfoSAM_b_leaf.sh`和`config/eval/InfoSAM_b_leaf.sh`在快照中缺失
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - 在打包材料中未检测到明确的API凭据或数据集访问令牌
+  - HuggingFace模型访问是公开的
+  - Facebook AI SAM检查点下载是公开的
+resources:
+  minimum:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 24
+  recommended:
+    cpu_cores: 32
+    ram_gb: 128
+    disk_gb: 200
+    gpu_count: 1
+    gpu_vram_gb: 48
+environment:
+  python: '3.8'
+  cuda: '12.1'
+  pytorch: 2.4.1
+  flash_attn: null
+  key_packages:
+  - torchvision==0.19.1
+  - transformers
+  - pillow
+  - numpy
+  - opencv-python
+  - scipy
+  notes:
+  - 测试环境：Ubuntu 20.04上的Python 3.8.20、PyTorch 2.4.1、torchvision 0.19.1、CUDA 12.1
+  - 环境规格来源于README文档；端到端执行需要恢复缺失的启动脚本
+risk_flags:
+- source_snapshot_incomplete
+- launch_scripts_missing
+- dataset_download_script_absent
+risk_notes:
+- 当前快照保留了文档（README、OPTIMIZATION.md、AGENTS.md、latex.md）但缺少可执行的训练/评估启动器树
+- 缺少数据集准备所需的`data/data_download.py`脚本
+- 缺少训练和评估所需的`config/train/InfoSAM_b_leaf.sh`和`config/eval/InfoSAM_b_leaf.sh`
+- 视为基于论文的重建目标；完整执行需要从官方仓库恢复上游脚本
+recommended_when: |
+  使用此基准来重建和审计InfoSAM PEFT方法论以实现SAM适配。适合研究信息论蒸馏用于视觉基础模型、关系知识迁移和领域无关分割的研究人员。请准备好从官方仓库恢复缺失的训练和评估脚本。
+not_recommended_when: |
+  不要用作可直接运行的本地基准。当前快照相对于官方README不完整。在基准可运行之前，需要从上游InfoSAM仓库恢复缺失的启动脚本、下载工具和配置文件。
+paper:
+  title: 'InfoSAM: Fine-Tuning the Segment Anything Model from An Information-Theoretic
+    Perspective'
+  venue: ICML 2025 Spotlight
+  year: 2025
+  url: https://proceedings.mlr.press/v267/zhang25ct.html
+  arxiv_url: https://arxiv.org/pdf/2505.21920
+  project_page: https://muyaoyuan.github.io/InfoSAM_Page
+  authors:
+  - Yuanhong Zhang
+  - Muyao Yuan
+  - Weizhan Zhang
+  - Tieliang Gong
+  - Wen Wen
+  - Jiangyong Ying
+  - Weijie Shi
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.056_infosam.zip
+  archive_type: zip
+  local_dir_name: paper-56-InfoSAM
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.056_infosam.zip
+  sha256: 6339252a6968df0ec70cdac20e1eb44e1eeae9f33b3b6386513a2772d1e51ba9
+  size_bytes: 45688
+display:
+  palette_seed: gold-forest-segment
+  art_style: info-geometry
+  accent_priority: high
+  domain_tags:
+  - computer_vision
+  - segmentation
+  - foundation_models
+  - distillation
+  - parameter_efficient_finetuning
+  - information_theory
+image_path: ../../../AISB/image/056_aisb.t3.056_infosam.jpg
+capability_tags:
+- research_code_optimization
+- computer_vision
+- segmentation
+- foundation_models
+- distillation
+- parameter_efficient_finetuning
+- mutual_information
+- knowledge_distillation
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+metric_contracts:
+  s_measure:
+    description: 用于前景分割质量评估的结构度量
+    origin: paper_evaluation
+    blocked_reason: 当前快照中无可执行的评估代码锚点
+  e_measure:
+    description: 用于边界和区域评估的增强对齐度量
+    origin: paper_evaluation
+    blocked_reason: 当前快照中无可执行的评估代码锚点
+  weighted_f_measure:
+    description: 结合精度和召回率并具有最优阈值选择的加权F度量
+    origin: paper_evaluation
+    blocked_reason: 当前快照中无可执行的评估代码锚点
+external_models:
+  teacher_models:
+  - model_id: sam_vit_b
+    url: https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth
+    description: ViT-Base SAM模型（默认教师）
+  - model_id: sam_vit_h
+    url: https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth
+    description: ViT-Huge SAM模型
+  - model_id: sam_vit_l
+    url: https://dl.fbaipublicfiles.com/segment_anything/sam_vit_l_0b3195.pth
+    description: ViT-Large SAM模型
+  pretrained_checkpoints:
+  - model_id: infosam_finetuned
+    url: https://huggingface.co/MuyaoYuan/InfoSAM/tree/main
+    description: HuggingFace上的预训练InfoSAM检查点
+    access: public
+restoration_required:
+- data/data_download.py
+- config/train/InfoSAM_b_leaf.sh
+- config/eval/InfoSAM_b_leaf.sh
+- requirements.txt
+- launcher scripts referenced in README
+restoration_notes: |
+  官方仓库包含完整的训练和评估脚本。重建工作涉及克隆上游仓库或恢复引用的shell脚本和下载工具，以实现完全可运行的基准。

package/AISB/catalog/aisb.t3.057_mdreid.yaml

@@ -0,0 +1,129 @@
+id: aisb.t3.057_mdreid
+name: MDReID
+version: 0.1.0
+one_line: Two-stage multi-modal object re-identification with modality-decoupled feature
+  learning for any-to-any query-gallery modality combinations.
+task_description: 'This packaged benchmark covers multi-modal object re-identification
+  with modality-decoupled feature handling and retrieval-time reranking. MDReID introduces
+  Modality Decoupling Learning (MDL) and Modality-aware Metric Learning (MML) to disentangle
+  modality-shared and modality-specific features, enabling effective retrieval under
+  both modality-matched and modality-mismatched query-gallery scenarios. Supports
+  three multi-spectral datasets: RGBNT201, RGBNT100, and MSVR310.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 6-24h
+cost_band: high
+difficulty: hard
+data_access: public
+primary_outputs:
+- map
+- rank1
+- reid_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: Run one packaged multi-modal re-identification test route on a single
+    dataset using pre-trained checkpoints.
+- id: full_train_eval
+  label: Full Train + Eval
+  description: Train the MDReID model from scratch and evaluate across all packaged
+    person/object re-identification datasets.
+dataset_download:
+  primary_method: mixed
+  sources:
+  - name: RGBNT201
+    url: https://drive.google.com/drive/folders/1EscBadX-wMAT56_It5lXY-S3-b5nK1wH
+    type: google_drive
+  - name: RGBNT100
+    url: https://pan.baidu.com/s/1xqqhN4Lctm3RcUdskG0Ug
+    type: baidu_pan
+    access_code: rjin
+  - name: MSVR310
+    url: https://drive.google.com/file/d/1IxI-fGiluPO_Ies6YjDHeTEuVYhFdYwD/view
+    type: google_drive
+  - name: CLIP ViT-B-16 Pretrained
+    url: https://pan.baidu.com/s/1YPhaL0YgpI-TQ_pSzXHRKw
+    type: baidu_pan
+    access_code: 52fu
+  notes:
+  - Download all person datasets before training; modify dataset paths in config files
+    for RGBNT100 and MSVR310
+  - Pretrained CLIP ViT-B-16 required for backbone initialization
+  - Pre-trained model checkpoints available for RGBNT201, RGBNT100, and MSVR310
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 16
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 300
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.8'
+  cuda: '11.3'
+  pytorch: 1.12.0
+  flash_attn: null
+  key_packages:
+  - torch-geometric==2.1.0
+  notes:
+  - CUDA 11.3 recommended; compatible with CUDA 11.x variants
+  - Install dependencies via requirements.txt
+  - Device ID configurable in config/defaults.py
+  - cudnn.benchmark enabled by default for training speed
+risk_flags:
+- large_dataset_storage
+- multi_source_download
+- baidu_pan_access_required
+risk_notes:
+- RGBNT100 and MSVR310 require Baidu Pan access with provided codes
+- Large multi-modal image datasets require significant disk space for training
+- No benchmark execution was performed during this packaging pass; runtime validation
+  recommended
+recommended_when: 'Use this benchmark when you need a multimodal retrieval task with
+  real person and object re-identification datasets under varying modality conditions.
+  Ideal for research on cross-modal retrieval, modality generalization, and vision-language
+  feature decoupling in surveillance or heterogeneous sensor scenarios.
+
+  '
+not_recommended_when: 'Do not use this if you cannot host large multi-modal image
+  datasets locally, if you need a text-centric or language-heavy benchmark, or if
+  you lack access to Baidu Pan for dataset downloads. Not recommended for single-GPU
+  memory below 16GB when training.
+
+  '
+paper:
+  title: 'MDReID: Modality-Decoupled Learning for Any-to-Any Multi-Modal Object Re-Identification'
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://openreview.net/forum?id=7jg26Fd1ra
+  code_url: https://github.com/stone96123/MDReID
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.057_mdreid.zip
+  archive_type: zip
+  local_dir_name: paper-57-MDReID
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.057_mdreid.zip
+  sha256: 443ff33a90ece8148f74b30103881c6835a07e69205d2d151c54fbc15215b9e8
+  size_bytes: 1554938
+display:
+  palette_seed: crimson-cobalt-reid
+  art_style: retrieval-vision
+  accent_priority: high
+image_path: ../image/057_aisb.t3.057_mdreid.jpg

package/AISB/catalog/aisb.t3.057_mdreid.zh.yaml

@@ -0,0 +1,117 @@
+id: aisb.t3.057_mdreid
+name: MDReID
+version: 0.1.0
+one_line: >-
+  两阶段多模态目标重识别，通过模态解耦特征学习支持任意到任意的查询-图库模态组合。
+task_description: >-
+  该封装基准测试涵盖多模态目标重识别任务，采用模态解耦特征处理和检索时重排序技术。
+  MDReID 提出了模态解耦学习（MDL）和模态感知度量学习（MML），将模态共享特征与模态特定特征分离，
+  使模态匹配和模态不匹配的查询-图库场景下均能实现有效检索。支持三个多光谱数据集：RGBNT201、RGBNT100 和 MSVR310。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 6-24h
+cost_band: high
+difficulty: hard
+data_access: public
+primary_outputs:
+- map
+- rank1
+- reid_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: >-
+    使用预训练权重在单一数据集上运行一条封装的多模态重识别测试流程。
+- id: full_train_eval
+  label: 完整训练+评估
+  description: >-
+    从头训练 MDReID 模型，并在所有封装的人员/物体重识别数据集上进行评估。
+dataset_download:
+  primary_method: mixed
+  sources:
+  - name: RGBNT201
+    url: https://drive.google.com/drive/folders/1EscBadX-wMAT56_It5lXY-S3-b5nK1wH
+    type: google_drive
+  - name: RGBNT100
+    url: https://pan.baidu.com/s/1xqqhN4Lctm3RcUdskG0Ug
+    type: baidu_pan
+    access_code: rjin
+  - name: MSVR310
+    url: https://drive.google.com/file/d/1IxI-fGiluPO_Ies6YjDHeTEuVYhFdYwD/view
+    type: google_drive
+  - name: CLIP ViT-B-16 Pretrained
+    url: https://pan.baidu.com/s/1YPhaL0YgpI-TQ_pSzXHRKw
+    type: baidu_pan
+    access_code: 52fu
+  notes:
+  - 训练前请下载所有人员数据集；如使用 RGBNT100 和 MSVR310，需修改配置文件中的数据集路径
+  - 预训练的 CLIP ViT-B-16 是骨干网络初始化所必需的
+  - RGBNT201、RGBNT100 和 MSVR310 均提供预训练模型权重
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 16
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 300
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.8'
+  cuda: '11.3'
+  pytorch: 1.12.0
+  flash_attn: null
+  key_packages:
+  - torch-geometric==2.1.0
+  notes:
+  - 推荐使用 CUDA 11.3；与 CUDA 11.x 系列兼容
+  - 通过 requirements.txt 安装依赖
+  - config/defaults.py 中可配置设备 ID
+  - 训练速度优化默认启用 cudnn.benchmark
+risk_flags:
+- large_dataset_storage
+- multi_source_download
+- baidu_pan_access_required
+risk_notes:
+- RGBNT100 和 MSVR310 需要使用提供的提取码访问百度云盘
+- 大型多模态图像数据集需要大量磁盘空间用于训练
+- 本次封装过程未执行基准测试；建议进行运行时验证
+recommended_when: >-
+  当您需要针对真实人员和物体重识别数据集在不同模态条件下进行多模态检索任务时使用此基准测试。
+  适用于监控或异构传感器场景下的跨模态检索、模态泛化以及视觉-语言特征解耦研究。
+not_recommended_when: >-
+  若您无法在本地托管大型多模态图像数据集，或需要以文本为中心、语言密集型的基准测试，
+  或无法访问百度云盘下载数据集，请勿使用。训练时单卡显存低于 16GB 不推荐使用。
+paper:
+  title: 'MDReID: Modality-Decoupled Learning for Any-to-Any Multi-Modal Object Re-Identification'
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://openreview.net/forum?id=7jg26Fd1ra
+  code_url: https://github.com/stone96123/MDReID
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.057_mdreid.zip
+  archive_type: zip
+  local_dir_name: paper-57-MDReID
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.057_mdreid.zip
+  sha256: 443ff33a90ece8148f74b30103881c6835a07e69205d2d151c54fbc15215b9e8
+  size_bytes: 1554938
+display:
+  palette_seed: crimson-cobalt-reid
+  art_style: retrieval-vision
+  accent_priority: high
+image_path: ../image/057_aisb.t3.057_mdreid.jpg