mudassiruddin/yolo

P2-Graft: runs, results and weights

When does a stride-4 prediction level help a pretrained YOLOv10 on drone images?

Quratulain Nayeem, Fahmina Taranum, Mohammed Mudassir Uddin, Sara Rizwan

This dataset repository holds everything behind the paper: per-run results, training curves, saved test predictions and the final weights of the 51 runs reported there. A stride-4 branch is grafted onto a pretrained YOLOv10 so that the original outputs stay unchanged at the start, and the new branch is initialised from the pretrained stride-8 weights.

Jump to: Results · Files · Use the weights · Reproduce

---

Results

AP on VisDrone test-dev (primary), validation AP in the paper. One recipe, one GPU, last-epoch weights, gains paired by seed.

Setting	Baseline	P2-Graft	Gain	Seeds
YOLOv10n, 640 px, 90 epochs	14.09	15.77	+1.68	3
YOLOv10n, 896 px, 90 epochs	18.16	18.74	+0.58	3
YOLOv10n, 1280 px, 90 epochs	22.45	22.32	−0.13	1
YOLOv10n, 640 px, 150 epochs	14.46	15.63	+1.17	1
YOLOv10s, 640 px, 90 epochs	17.46	18.19	+0.73	3

The gain is largest at 640 px and shrinks as the input grows; at equal GFLOPs a plain baseline at higher resolution is more accurate. On YOLOv10s the per-seed gains are +0.06, +0.19 and +1.94.

Validation AP against published results (ours in bold)

Model	Input	Params (M)	AP	AP50
YOLOv8-N	640	3.2	19.6	33.3
FBRT-YOLO-N	640	0.9	20.2	34.4
YOLOv10-S	640	7.2	23.8	39.3
FBRT-YOLO-S	640	2.9	25.9	42.4
DroneScan-YOLO	1280	10.23	35.6	55.3
P2-Graft, YOLOv10n	640	2.41	20.64	35.96
P2-Graft, YOLOv10s	640	7.58	23.29	39.86
YOLOv10n baseline	1536	2.27	30.29	49.42

Published numbers are quoted from FBRT-YOLO and DroneScan-YOLO with their own schedules; the rows in bold are 90-epoch runs.

What matters inside the branch (35 epochs, 640 px)

The warm start, copying the pretrained stride-8 block and head into the new level, is what makes the branch useful: +0.92 test-dev AP against +0.26 for a random start.

Qualitative examples (validation crops)

Green: match, red: false alarm, yellow: missed. The two crops show the largest gain in found objects.

Files

README.md                         this card
assets/                           figures of the paper
paper/main.pdf                    the paper
results/                          runs_summary.csv (the 51 runs), final_runs.csv (all metrics), contrasts_session_{A,B,C}.csv
sessionA-20261104-0048/           896 px, 90 epochs, 3 seeds: model and baseline
sessionB-20261106-1639/           1280 px, 150-epoch schedule, batch-32 control
sessionC-20261107-0846/           YOLOv10s, 3 seeds
  runs/<run id>/                  result.json, results.csv (training curve), args.yaml,
                                  weights/last.pt (final weights), preds_*.npz (saved predictions)
  results/                        leaderboard.csv, contrasts.csv (image-bootstrap intervals), all_runs.csv
  logs/                           worker and resume logs

Run ids read <family>-<variant>-<tag>-<input size>-e<epochs>-s<seed>: variant base is the unmodified YOLOv10, graftd with transfer role_ws is P2-Graft (deeper block, warm start). Runs of the 35-epoch ablation and of the 640 px comparisons come from earlier sessions and are listed with their source in results/final_runs.csv.

Use the weights

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

path = hf_hub_download(
    "mudassiruddin/yolo", repo_type="dataset",
    filename="sessionB-20261106-1639/runs/v10-graftd-ws-640-e150-s0/weights/last.pt")
model = YOLO(path)                       # P2-Graft, YOLOv10n, 640 px, 150 epochs
model.predict("drone_image.jpg", imgsz=640, conf=0.25)

The checkpoints store the four-level architecture, so Ultralytics rebuilds it from the file.

Reproduce

Code, notebook and the scripts behind every table and figure: github.com/Mudassiruddin7/YOLO. All runs used one NVIDIA RTX PRO 6000 Blackwell GPU, PyTorch 2.11 and Ultralytics 8.4.150.

Licence and data

The weights derive from Ultralytics YOLOv10 checkpoints and keep the AGPL-3.0 licence. VisDrone images and annotations are not included; they belong to the VisDrone project.