ReID with EfficientNet-V2 and Augmentations

A Person Re-Identification (ReID) project built with PyTorch Lightning that explores the use of EfficientNet-V2 models and specialized ReID augmentations to improve identification performance.

Overview

This project implements a person re-identification system with the following key features:

• Multiple backbone architectures: ResNet, ResNet-IBN-A, EfficientNet, and EfficientNet-V2
• ReID-specific augmentations: Local Grayscale Transformation (LGT), Local Grayscale Patch Replacement (LGPR), Global Grayscale Patch Replacement (GGPR), and RGB-Gray-Sketch fusion
• Centroid-based evaluation: Optional use of centroids during validation for improved performance
• Mixed precision training: Support for faster training with automatic mixed precision

Requirements

Install the required dependencies:

bash pip install -r requirements.txt

Key dependencies include: - PyTorch Lightning 1.1.4 - PyTorch 1.7.1+cu101 - EfficientNet PyTorch - YACS for configuration management

Dataset Setup

1. Prepare your dataset in COCO format with the following structure: dataset/ ├── images/ │ ├── train/ │ ├── query/ │ └── gallery/ └── annotations/ ├── train_reid.json ├── query_reid.json └── gallery_reid.json

2. For DeepFashion Consumer-to-Shop dataset, use the conversion script: bash python scripts/deep_fashion2reid.py --root-dir-path /path/to/deepfashion --target-image-size 320 320

Configuration

Create or modify configuration files in the configs/ directory. Example configurations are provided:

• configs/256_resnet50.yml - ResNet50 with 256x128 input
• configs/320_efficientnet_b0.yml - EfficientNet-B0 with 320x320 input
• configs/320_efficientnet_b1.yml - EfficientNet-B1 with 320x320 input

Key Configuration Options

```yaml MODEL: NAME: 'efficientnet-b0' # Model architecture USE_CENTROIDS: True # Use centroids for evaluation

INPUT: SIZETRAIN: [320, 320] # Training image size USELGT: False # Local Grayscale Transformation USELGPR: True # Local Grayscale Patch Replacement
USEGGPR: True # Global Grayscale Patch Replacement USEFUSERGB: False # RGB-Gray-Sketch fusion

SOLVER: BASELR: 0.0001 # Learning rate MAXEPOCHS: 50 # Training epochs IMSPERBATCH: 32 # Batch size

DATASETS: NAMES: 'df1' ROOTDIR: '/path/to/dataset' JSONTRAINPATH: '/path/to/trainannotations.json' ```

Training

Create a Training Script

Create a training script (e.g., train_efficientnet_v2.sh):

```bash

!/bin/bash

python trainctlmodel.py \ --configfile="configs/320efficientnetb0.yml" \ MODEL.NAME 'efficientnet-b0' \ GPUIDS [0] \ DATASETS.NAMES 'df1' \ DATASETS.JSONTRAINPATH 'dataset/trainreidcropped320320.json' \ DATASETS.ROOTDIR 'dataset/320320images' \ SOLVER.IMSPERBATCH 32 \ TEST.IMSPERBATCH 32 \ SOLVER.BASELR 1e-4 \ SOLVER.MAXEPOCHS 50 \ OUTPUTDIR './logs/efficientnetb0augmented' \ DATALOADER.USERESAMPLING False \ MODEL.KEEPCAMIDCENTROIDS False \ INPUT.USELGT False \ INPUT.USELGPR True \ INPUT.USEGGPR True \ INPUT.USEFUSERGB True \ MODEL.USECENTROIDS True \ EXPERIMENTNAME 'efficientnetb0with_augmentations' ```

Run Training

bash chmod +x train_efficientnet_v2.sh ./train_efficientnet_v2.sh

Or run directly: bash python train_ctl_model.py --config_file="configs/320_efficientnet_b0.yml" [additional parameters]

Project Novelty

This project introduces several key innovations:

1. EfficientNet-V2 Integration: First implementation using EfficientNet-V2 models for person ReID, providing better efficiency and accuracy trade-offs.

2. ReID-Specific Augmentations:

   • LGT (Local Grayscale Transformation): Randomly converts local patches to grayscale
   • LGPR (Local Grayscale Patch Replacement): Replaces random patches with grayscale versions
   • GGPR (Global Grayscale Patch Replacement): Global grayscale transformations
   • RGB-Gray-Sketch Fusion: Combines RGB, grayscale, and sketch representations

3. Comparative Analysis: Systematic evaluation of augmentation effects on different backbone architectures.

Evaluation

The model automatically evaluates on validation data during training. Key metrics include: - mAP (mean Average Precision): Primary evaluation metric - CMC (Cumulative Matching Characteristics): Rank-1, Rank-5, Rank-10 accuracy - Top-K accuracy: For various K values

Model Checkpoints

Models are automatically saved during training: - Best model based on mAP is saved automatically - Periodic checkpoints every N epochs (configurable) - Checkpoints include full model state for resuming training

Testing Only

To run evaluation on a trained model:

bash python train_ctl_model.py \ --config_file="configs/your_config.yml" \ TEST.ONLY_TEST True \ TEST.WEIGHT "/path/to/checkpoint.pth" \ [other parameters]

Logs and Monitoring

Training logs and metrics are automatically saved to: - TensorBoard logs in the output directory - MLflow tracking (if configured) - Console output with detailed metrics

Tips for Best Results

1. Batch Size: Adjust based on GPU memory (32 works well for most setups)
2. Learning Rate: Start with 1e-4, adjust based on convergence
3. Augmentations: Try different combinations of LGPR, GGPR, and RGB fusion
4. Image Size: 320x320 generally works better than 256x128 for EfficientNet models
5. Centroids: Enable USE_CENTROIDS for better evaluation performance

Example Training Commands

EfficientNet-B0 with All Augmentations

bash python train_ctl_model.py \ --config_file="configs/320_efficientnet_b0.yml" \ MODEL.NAME 'efficientnet-b0' \ GPU_IDS [0] \ DATASETS.NAMES 'df1' \ DATASETS.JSON_TRAIN_PATH 'dataset/train_reid_cropped_320_320.json' \ DATASETS.ROOT_DIR 'dataset/320_320_images' \ SOLVER.IMS_PER_BATCH 32 \ TEST.IMS_PER_BATCH 32 \ SOLVER.BASE_LR 1e-4 \ SOLVER.MAX_EPOCHS 50 \ OUTPUT_DIR './logs/efficientnet_b0_all_aug' \ INPUT.USE_LGPR True \ INPUT.USE_GGPR True \ INPUT.USE_FUSE_RGB True \ MODEL.USE_CENTROIDS True \ EXPERIMENT_NAME 'efficientnet_b0_full_augmentation'

EfficientNet-V2-S with Selective Augmentations

bash python train_ctl_model.py \ --config_file="configs/320_efficientnet_v2_s.yml" \ MODEL.NAME 'efficientnet_v2-s' \ GPU_IDS [0] \ DATASETS.NAMES 'df1' \ DATASETS.JSON_TRAIN_PATH 'dataset/train_reid_cropped_320_320.json' \ DATASETS.ROOT_DIR 'dataset/320_320_images' \ SOLVER.IMS_PER_BATCH 24 \ TEST.IMS_PER_BATCH 24 \ SOLVER.BASE_LR 1e-4 \ SOLVER.MAX_EPOCHS 50 \ OUTPUT_DIR './logs/efficientnet_v2_s_selective' \ INPUT.USE_LGPR True \ INPUT.USE_GGPR False \ INPUT.USE_FUSE_RGB False \ MODEL.USE_CENTROIDS True \ EXPERIMENT_NAME 'efficientnet_v2_s_lgpr_only'

ResNet50 Baseline

bash python train_ctl_model.py \ --config_file="configs/256_resnet50.yml" \ MODEL.NAME 'resnet50' \ GPU_IDS [0] \ DATASETS.NAMES 'df1' \ DATASETS.JSON_TRAIN_PATH 'dataset/train_reid_cropped_320_320.json' \ DATASETS.ROOT_DIR 'dataset/320_320_images' \ SOLVER.IMS_PER_BATCH 32 \ TEST.IMS_PER_BATCH 32 \ SOLVER.BASE_LR 1e-4 \ SOLVER.MAX_EPOCHS 50 \ OUTPUT_DIR './logs/resnet50_baseline' \ INPUT.USE_LGPR False \ INPUT.USE_GGPR False \ INPUT.USE_FUSE_RGB False \ MODEL.USE_CENTROIDS True \ EXPERIMENT_NAME 'resnet50_baseline'