Skip to content

README.md

Latest commit

 

History

History
160 lines (117 loc) · 6.99 KB

README.md

File metadata and controls

160 lines (117 loc) · 6.99 KB

Small Objects Detection Model Zoo

Link to models stored on Kaggle

This folder contains trained detectors used in the small objects detection benchmark trained on the SkyFusion dataset. These files are intended for use in Kaggle notebooks or locally. On Kaggle, place the files in your working directory or reference them from your Dataset path (e.g., /kaggle/input/small-objects-detection-benchmark/models/<model-name>.pt).

Final models

Name Benchmark label Type File Framework Optimal params
yolov8m-aug-update_20250603 YOLOv8m (Optimized Aug, CosLR) yolo yolov8m-aug-update_20250603.pt Ultralytics (YOLOv8) conf_thr=0.06291186979342091, nms_iou=0.3358295483691517
yolov11m-aug_20250603 YOLOv11m (Optimized Aug) yolo yolov11m-aug_20250603.pt Ultralytics (YOLOv11) N/A
yolov11m-p2-aug_20250603 YOLOv11m (P2, Optimized Aug) yolo yolov11m-p2-aug_20250603.pt Ultralytics (YOLOv11) conf_thr=0.052566007120515956, nms_iou=0.49317179138811856
rf-detr RF-DETR rfdetr rfdetr_best_total.pth PyTorch (RF-DETR impl) conf_thr=0.09616820140192325
faster-rcnn Faster R-CNN faster-rcnn fasterrcnn-best-epoch=18-val_map=0.31.ckpt PyTorch Lightning + torchvision conf_thr=0.07957236023833904, nms_iou=0.621230971215935
rt-detr RT-DETR-L rtdetr rtdetr-aug_best.pt Ultralytics (RT-DETR) conf_thr=0.2704984199324548

Notes:

  • The table includes the exact optimal thresholds; you don’t need an external YAML file.
  • If a model doesn’t list nms_iou, use the framework’s default or your benchmark’s setting.

YOLO training progression (changes and artifacts)

This table summarizes the step-by-step progression used in experiments and links steps to any published artifacts.

Step Progression ID Key change Artifact file Benchmark label
1 Baseline Testing Size sweep — —
2 yolov8m-baseline Standard training yolov8m-baseline_20250518.pt YOLOv8m (Baseline)
3 yolov8m-baseline-aug Add augmentation yolov8m-baseline-aug_20250518.pt YOLOv8m (Aug)
4 yolov8m-cos-lr Cosine learning rate yolov8m-aug-100epochs-cos-lr_20250518.pt YOLOv8m (Aug, CosLR)
5 yolov8m-ships-vehicles Exclude aircraft (failed) yolov8m-aug-100epochs-ships-vehicles_20250518.pt —
6 yolov8m-aug-update Optimize augmentations yolov8m-aug-update_20250603.pt YOLOv8m (Optimized Aug, CosLR)
7 yolov11m-aug Try YOLOv11 yolov11m-aug_20250603.pt YOLOv11m (Optimized Aug)
8 yolov11m-p2-aug Add P2 enhancement yolov11m-p2-aug_20250603.pt YOLOv11m (P2, Optimized Aug)

Notes:

  • “Benchmark label” matches how models are referred to in result tables.

Ultralytics (YOLOv8 / YOLOv11)

Install on Kaggle:

pip install -q ultralytics

YOLOv11 example:

from ultralytics import YOLO

model = YOLO("yolov11m-p2-aug_20250603.pt")
conf_thr = 0.052566007120515956
nms_iou  = 0.49317179138811856

results = model("image.jpg", conf=conf_thr, iou=nms_iou)
# results[0].plot()  # visualize if needed

YOLOv8 example:

from ultralytics import YOLO

model = YOLO("yolov8m-aug-update_20250603.pt")
conf_thr = 0.06291186979342091
nms_iou  = 0.3358295483691517

results = model("image.jpg", conf=conf_thr, iou=nms_iou)

Ultralytics RT-DETR

# Depending on ultralytics version:
from ultralytics import RTDETR, YOLO

# Preferred:
model = RTDETR("rtdetr-aug_best.pt")

# Fallback (unified loader):
# model = YOLO("rtdetr-aug_best.pt")

conf_thr = 0.2704984199324548
results = model("image.jpg", conf=conf_thr)  # iou optional for RT-DETR

Faster R-CNN (PyTorch Lightning)

The checkpoint is a Lightning .ckpt. Replace MyFasterRCNNModule with your module class.

pip install -q lightning
import torch
from lightning import LightningModule

class MyFasterRCNNModule(LightningModule):
    def __init__(self, ...):
        super().__init__()
        # define model here (e.g., torchvision.models.detection.fasterrcnn_resnet50_fpn)
        ...

ckpt_path = "fasterrcnn-best-epoch=18-val_map=0.31.ckpt"
model = MyFasterRCNNModule.load_from_checkpoint(ckpt_path, strict=False)
model.eval()

# Implement your preprocessing/postprocessing for inference
# outputs = model(images)

If .ckpt stores a plain state_dict:

import torch

state = torch.load("fasterrcnn-best-epoch=18-val_map=0.31.ckpt", map_location="cpu")
model = build_faster_rcnn_model()  # your builder returning an nn.Module
model.load_state_dict(state.get("state_dict", state), strict=False)
model.eval()

Recommended inference thresholds:

  • conf_thr: 0.07957236023833904
  • nms_iou: 0.621230971215935

RF-DETR (PyTorch)

pip install -q "rfdetr[metrics]"
from rfdetr import RFDETRBase
from PIL import Image

# Initialize with pretrained weights at construction (matches adapter)
model = RFDETRBase(
    device="cuda",  # or "cpu"
    pretrain_weights="rfdetr_best_total.pth",
)

# Inference on a single image using the optimal threshold
threshold = 0.09616820140192325  # optimal conf_thr
pred = model.predict("image.jpg", threshold=threshold)

# Alternatively, pass a PIL image (adapter converts BGR->RGB then uses PIL)
# img = Image.open("image.jpg").convert("RGB")
# pred = model.predict(img, threshold=threshold)