Team C-B members

Name	GitHub	LinkedIn	Zindi
Durojaiye Abisoye	https://github.com/DurojaiyeAbisoye	www.linkedin.com/in/abisoye-durojaiye	https://zindi.africa/users/Bisoye
Agbaje Ayomipo	https://github.com/AgbajeAyomipo	https://www.linkedin.com/in/agbaje-ayomipo-6b742423b	https://zindi.africa/users/crossentropy

Overview and Objectives

This solution aims to develop an object detection model to identify multiple diseases in cocoa plant images. Our solution is designed to assist subsistence farmers in Africa by enabling disease detection using entry-level smartphones. The objectives include:

• Accurate multi-class disease detection on cocoa plant images.
• Generalization to unseen diseases not present in the training set.
• Efficient deployment and inference on edge devices.

Folder structure

├── README.md  
├── requirements.txt  
├── notebooks/  
│   ├── data-preparation.ipynb  
│   ├── rfdetr-training.ipynb  
│   ├── rfdetr-inference.ipynb  
│   └── wbf.ipynb  
└── results/  
    ├── inference output.csv  
    └── rfdetr_9_epochs_full_linear_best_ema_wbf-max.csv  

Approach

• Model: We make use of rfdetr, a new SOTA real-time object detection model released by roboflow, that supports ONNX deployment natively.
• Data preparation(data-preparation.ipynb):
  • We split the dataset into train and validation sets using StratifiedGroupKFold with k of 10. We use fold 0 for validation and the other folds for training.
  • We convert our split dataset into COCO format as required by the rfdetr

dataset/
├── train/
│   ├── _annotations.coco.json
│   ├── image1.jpg
│   ├── image2.jpg
│   └── ... (other image files)
├── valid/
│   ├── _annotations.coco.json
│   ├── image1.jpg
│   ├── image2.jpg
│   └── ... (other image files)
└── test/
   ├── _annotations.coco.json
   ├── image1.jpg
   ├── image2.jpg
   └── ... (other image files)

• Training(rfdetr-training.ipynb): We trained the RfDetr large for 9 epochs using the following configuration
  • epochs: 9
  • batch size: 2
  • grad_accum_steps: 8
  • lr: 1e-4
  • resolution: 448
• Inference(rfdetr-inference.ipynb): RfDetr produces 2 checkpoints: regular weights and EMA-based weights. We made use of the EMA-based weights as those have better performance locally and on the leaderboard. We used a low confidence threshold of 0.02
• Ensembling(wbf.ipynb): We use wbf on the output from the inference notebook. We made use of the ensemble-boxes framework with the following parameters:
  • conf_type: 'max'
  • skip_box_thr: 0.0001
  • iou_thr: 0.6

You can find the full training hparams here:

Namespace(num_classes=3, grad_accum_steps=8, amp=True, lr=0.0001, lr_encoder=0.00015, batch_size=2,
weight_decay=0.0001, epochs=9, lr_drop=100, clip_max_norm=0.1, lr_vit_layer_decay=0.8, lr_component_decay=0.7,
do_benchmark=False, dropout=0, drop_path=0.0, drop_mode='standard', drop_schedule='constant', cutoff_epoch=0,
pretrained_encoder=None, pretrain_weights='rf-detr-large.pth', pretrain_exclude_keys=None, pretrain_keys_modify_to_load=None, pretrained_distiller=None, encoder='dinov2_windowed_base', vit_encoder_num_layers=12, window_block_indexes=None,
position_embedding='sine', out_feature_indexes=[2, 5, 8, 11], freeze_encoder=False, layer_norm=True, rms_norm=False,
backbone_lora=False, force_no_pretrain=False, dec_layers=3, dim_feedforward=2048, hidden_dim=384, sa_nheads=12, ca_nheads=24, num_queries=300, group_detr=13, two_stage=True, projector_scale=['P3', 'P5'], lite_refpoint_refine=True, num_select=300,
dec_n_points=4, decoder_norm='LN', bbox_reparam=True, freeze_batch_norm=False, set_cost_class=2, set_cost_bbox=5, set_cost_giou=2,
cls_loss_coef=1.0, bbox_loss_coef=5, giou_loss_coef=2, focal_alpha=0.25, aux_loss=True, sum_group_losses=False,
use_varifocal_loss=False, use_position_supervised_loss=False, ia_bce_loss=True, dataset_file='roboflow', coco_path=None,
dataset_dir='/kaggle/input/amini-cocoa-disease-coco-dataset/amini cocoa disease coco dataset', square_resize_div_64=True,
output_dir='output', dont_save_weights=False, checkpoint_interval=10, seed=42, resume='', start_epoch=0, eval=False, use_ema=True,
ema_decay=0.993, ema_tau=100, num_workers=2, device='cuda', world_size=1, dist_url='env://', sync_bn=True, fp16_eval=False, encoder_only=False, backbone_only=False, resolution=448, use_cls_token=False, multi_scale=True, expanded_scales=True, warmup_epochs=0,
lr_scheduler='step', lr_min_factor=0.0, early_stopping=False, early_stopping_patience=10, early_stopping_min_delta=0.001,
early_stopping_use_ema=False, gradient_checkpointing=False, tensorboard=True, wandb=True, project='Amini cocoa rf-detr',
run='9 epochs from beginning', class_names=[], distributed=False)

Metrics

• Train Loss: 6.01317
• Test Loss: 6.74589
• Base Model AP metrics
  • AP50: 0.78086
  • AP50_90: 0.54817
  • AR50_90: 0.43859
• EMA Model AP metrics
  • AP50: 0.80309
  • AP50_90: 0.57591
  • AR50_90: 0.44415

Runtime

• Training: 8h 24m
• Inference: 21m

Submission Scores:

• Public Leaderboard: 0.825360226
• Private Leaderboard: 0.825496953