README_ARCHITECTURE.md

Varroa Object Detection Notes

Data

Repo data is already split as:

train|val|test/
  videos/<video-id>/*.png
  labels/<video-id>/*.txt

Each label file starts with an object count, then one or more boxes in absolute pixel x1 y1 x2 y2 format. Class 0 is varroa.

All current images are 160x280 pixels. The custom LMNet detector defaults to input_width=160, input_height=288, so it preserves width and only pads height to a stride-16 friendly size. YOLO baselines can still use larger imgsz values because Ultralytics handles resizing internally.

YOLOv8 Breakdown

YOLOv8 detection is usually organized as:

image
  -> Backbone: Conv/C2f/SPPF feature extractor
  -> Neck: PAN-FPN feature fusion, typically P3/P4/P5
  -> Head: decoupled anchor-free box/class branches
  -> Loss: bbox regression + classification + DFL

For this dataset, the useful hook points are:

• Backbone replacement: try LMNet encoder features instead of YOLO C2f stages.
• Neck replacement: keep PAN/FPN behavior because Varroa is small and multi-scale fusion matters.
• Head replacement: easiest to test outside Ultralytics first with a small anchor-free head.

YOLOv10 Breakdown

YOLOv10 keeps the YOLO-style backbone/neck/head split, but its training/inference design focuses on NMS-free detection via dual assignment:

image
  -> Backbone: efficient CSP/C2f-style stages
  -> Neck: PAN-FPN
  -> Head: one-to-many branch for training signal + one-to-one branch for final prediction
  -> Inference: designed to avoid NMS

This makes YOLOv10 less convenient to partially rewire unless you are editing the detector internals. Use it first as a baseline through Ultralytics, then compare with the standalone LMNet detector.

LMNet Detector In This Folder

Implemented path:

image
  -> LMNetBackbone
       stem:   stride 1
       stage2: stride 2
       stage3: stride 4  -> P3
       stage4: stride 8  -> P4
       p5:     stride 16 -> P5, optional PyramidPool + GFT
       optional NATTEN local attention on P3/P4/P5
  -> YOLOPANNeck
       upsample + concat top-down
       downsample + concat bottom-up
  -> FCOSHead
       class logit
       l/t/r/b box distances
       centerness
  -> FCOSLoss

Files:

• architectures/lmnet_backbone.py: LMNet encoder adapted from the segmentation model.
• architectures/common.py: FPN/PAN neck blocks and feature location helpers.
• architectures/fcos_head.py: detection head.
• architectures/lmnet_fcos.py: full model and prediction decode.
• losses.py: FCOS-style assignment/loss.
• metrics.py: IoU, NMS, precision/recall/F1.
• data/varroa_detection_dataset.py: direct loader for current data layout.

Commands

Prepare YOLO-format data:

python object_detection_related/prepare_yolo_dataset.py --root . --out-dir yolo_varroa_dataset

Train YOLOv8 baseline:

python object_detection_related/train_yolo_varroa.py train \
  --root . \
  --weights yolov8n.pt \
  --epochs 100 \
  --imgsz 640 \
  --batch-size 8 \
  --name yolov8n_varroa

Train YOLOv10 baseline, if your Ultralytics install has YOLOv10 weights:

python object_detection_related/train_yolo_varroa.py train \
  --root . \
  --weights yolov10n.pt \
  --epochs 100 \
  --imgsz 640 \
  --batch-size 8 \
  --name yolov10n_varroa

Evaluate a YOLO checkpoint:

python object_detection_related/train_yolo_varroa.py eval \
  --root . \
  --split test \
  --weights runs_yolo_varroa/yolov8n_varroa/weights/best.pt

Evaluation logs fixed-threshold precision/recall/F1 at --conf and COCO-style single-class mAP50 / mAP50_95. AP candidates are collected with --map-conf 0.001 by default, then fixed-threshold metrics are filtered back to --conf.

Train LMNet + detection head:

python object_detection_related/train_lmnet_detector_varroa.py \
  --root . \
  --input-height 288 \
  --input-width 160 \
  --batch-size 4 \
  --epochs 100 \
  --gft-kind global \
  --gft-bottleneck 128 \
  --checkpoint checkpoints/lmnet_fcos_varroa_best.pt

Train LMNet + YOLO-style PAN neck + NATTEN local attention:

python object_detection_related/train_lmnet_detector_varroa.py \
  --root . \
  --input-height 288 \
  --input-width 160 \
  --batch-size 4 \
  --epochs 100 \
  --neck pan \
  --local-attention natten \
  --local-attention-kernel 3 \
  --gft-kind global \
  --gft-bottleneck 128 \
  --checkpoint checkpoints/lmnet_pan_natten_varroa_best.pt

Cheap ablation without PyramidPool/GFT:

python object_detection_related/train_lmnet_detector_varroa.py \
  --root . \
  --input-height 288 \
  --input-width 160 \
  --batch-size 4 \
  --epochs 100 \
  --neck pan \
  --no-pyramid \
  --local-attention none

Evaluate LMNet + detection head:

python object_detection_related/eval_lmnet_detector_varroa.py \
  --root . \
  --split test \
  --checkpoint checkpoints/lmnet_fcos_varroa_best.pt

Training CSV columns include val_map50, val_map50_95, val_ap50, val_ap55, ..., val_ap95. The best checkpoint defaults to --best-metric map50_95.

Suggested Experiment Order

1. yolov8n.pt baseline.
2. yolov10n.pt baseline if available in your environment.
3. LMNetBackbone + YOLOPANNeck + FCOSHead with --no-pyramid --local-attention none.
4. LMNetBackbone + YOLOPANNeck + FCOSHead with --gft-kind global.
5. Add --local-attention natten after the non-attention variant trains cleanly.
6. Compare F1/precision/recall on the same test split and same confidence threshold.