README1.md

Tiny Object Dectection-Benchmarks-FullImage Vs GOIS

BY MUZAMMUL(ZJU)

Enhancing Tiny Object Detection Without Fine-Tuning: Dynamic Adaptive Guided Object Inference Slicing Framework with Latest YOLO Models and RT-DETR Transformer

Here is a demonstration video for the project:

(https://youtu.be/T5t5eb_w0S4)

Section1: WITHOUT FINE TUNING 15% DATASET SUBSET VISDRONE2019Train DATASET

Quick Steps for COCO Evaluation

This benchmark evaluates Full Image Inference Detection vs. ** Guided-Object Inference Slicing (GOIS)** for small object detection using COCO metrics. Results are based on a subset of 970 images (15%) from the VisDrone-2019-Train dataset.

1. Clone the Repository

Clone this repository to access the required files:[clone] (https://github.com/MMUZAMMUL/Small-Object-Detection-Benchmarks-Full_ImageVsGOIS.git) cd Small-Object-Detection-Benchmarks-Full_ImageVsGOIS

Instructions:

1. pip install pycocotools .
2. Open evaluation.py .
3. Update the paths: . Ground Truth Path: Set the path for Ground_Truth & Prediction Path: Download the prediction .json file for either Full Image or GOIS benchmarks and set its path..
4. Run python evaluation.py

---

Full Image Inference Results with Seven Different Models

Model	Predictions-Link	[email protected]:0.95	[email protected]	[email protected]	mAP-Small	mAP-Medium	mAP-Large	AR@1	AR@10	AR@100	AR-Small	AR-Medium	AR-Large	F1 Score
YOLO11	Download	0.12	0.18	0.13	0.02	0.23	0.57	0.12	0.27	0.29	0.04	0.49	1.09	0.17
RT-DETR-L	Download	0.43	0.67	0.44	0.11	0.67	1.34	0.32	0.81	1.01	0.44	1.44	2.45	0.61
YOLOv10	Download	0.12	0.17	0.13	0.02	0.18	0.63	0.13	0.25	0.27	0.02	0.38	1.18	0.17
YOLOv9	Download	0.41	0.56	0.45	0.06	0.72	1.33	0.30	0.65	0.73	0.17	1.20	2.22	0.52
YOLOv8n	Download	0.14	0.20	0.14	0.03	0.24	0.54	0.15	0.29	0.32	0.04	0.50	1.22	0.19
YOLOv5n	Download	0.18	0.27	0.19	0.03	0.32	0.79	0.16	0.36	0.41	0.10	0.67	1.51	0.25
YOLOv8s-WorldV2	Download	0.23	0.34	0.23	0.04	0.42	0.90	0.21	0.42	0.46	0.11	0.75	1.79	0.30

Guided-Object Inference Slicing (GOIS) Results with Seven Different Models

Model	Predictions-Link	[email protected]:0.95	[email protected]	[email protected]	mAP-Small	mAP-Medium	mAP-Large	AR@1	AR@10	AR@100	AR-Small	AR-Medium	AR-Large	F1 Score
YOLO11	Download	0.33	0.51	0.34	0.10	0.57	0.96	0.27	0.68	0.87	0.33	1.40	1.93	0.47
RT-DETR-L	Download	0.61	0.94	0.63	0.22	0.95	1.49	0.46	1.16	1.71	1.03	2.25	2.73	0.90
YOLOv10	Download	0.31	0.48	0.33	0.08	0.56	0.93	0.26	0.61	0.76	0.27	1.25	1.85	0.44
YOLOv9	Download	0.53	0.76	0.58	0.18	0.90	1.18	0.40	0.91	1.16	0.53	1.79	2.21	0.73
YOLOv8n	Download	0.30	0.47	0.32	0.13	0.53	0.97	0.28	0.67	0.84	0.39	1.34	1.93	0.44
YOLOv5n	Download	0.38	0.58	0.41	0.16	0.65	1.02	0.29	0.71	0.93	0.51	1.44	1.93	0.54
YOLOv8s-WorldV2	Download	0.40	0.60	0.43	0.16	0.68	1.01	0.36	0.84	1.03	0.48	1.59	1.97	0.58

Percentage(%age) improvement achived by GOIS over Full image

---

Model	[email protected]:0.95 %↑	[email protected] %↑	[email protected] %↑	mAP-Small %↑	mAP-Medium %↑	mAP-Large %↑	AR@1 %↑	AR@10 %↑	AR@100 %↑	AR-Small %↑	AR-Medium %↑	AR-Large %↑	F1 Score %↑
YOLO11	175.0	183.33	161.54	400.0	147.83	68.42	125.0	151.85	200.0	725.0	185.71	77.06	176.47
RT-DETR-L	41.86	40.3	43.18	100.0	41.79	11.19	43.75	43.21	69.31	134.09	56.25	11.43	47.54
YOLOv10	158.33	182.35	153.85	300.0	211.11	47.62	100.0	144.0	181.48	1250.0	228.95	56.78	158.82
YOLOv9	29.27	35.71	28.89	200.0	25.0	-11.28	33.33	40.0	58.9	211.76	49.17	-0.45	40.38
YOLOv8n	114.29	135.0	128.57	333.33	120.83	79.63	86.67	131.03	162.5	875.0	168.0	58.2	131.58
YOLOv5n	111.11	114.81	115.79	433.33	103.12	29.11	81.25	97.22	126.83	410.0	114.93	27.81	116.0
YOLOv8s-WorldV2	73.91	76.47	86.96	300.0	61.9	12.22	71.43	100.0	123.91	336.36	112.0	10.06	93.33

Section 2: Fine Tuning Full Dataset Subset VisDrone2019Train Dataset

Comparative Results for FI-Det and GOIS-Det

This table presents the Average Precision (AP) and Average Recall (AR) metrics for five models (YOLO11, YOLOv10, YOLOv9, YOLOv8, YOLOv5). Each model includes three rows: FI-Det results, GOIS-Det results, and % improvement achieved by GOIS. Downloadable links for FI-Det and GOIS-Det results are included in the first column next to the model name.

Model	FI-Det File	GOIS-Det File	AP@[IoU=0.50:0.95]	AP@[IoU=0.50]	AP@[IoU=0.75]	AP-Small	AP-Medium	AP-Large	AR@1	AR@10	AR@100	AR-Small	AR-Medium	AR-Large	F1 Score
YOLO11	Download	Download	0.12	0.171	0.119	0.024	0.159	0.283	0.045	0.112	0.137	0.035	0.208	0.349	0.17
			0.123	0.193	0.132	0.072	0.164	0.151	0.054	0.152	0.208	0.132	0.274	0.227	0.47
			↑ 12.01%	↑ 12.38%	↑ 11.26%	↑ 196.90%	↑ 2.94%	↓ 46.71%	↑ 18.81%	↑ 35.46%	↑ 51.17%	↑ 278.66%	↑ 31.44%	↓ 34.90%	↑ 176.47%
YOLOv10	Download	Download	0.091	0.140	0.100	0.022	0.133	0.222	0.041	0.097	0.117	0.029	0.178	0.278	0.17
			0.099	0.156	0.107	0.061	0.130	0.100	0.047	0.127	0.172	0.109	0.219	0.159	0.44
			↑ 8.88%	↑ 11.40%	↑ 7.08%	↑ 176.54%	↓ 2.30%	↓ 54.85%	↑ 14.18%	↑ 31.01%	↑ 46.09%	↑ 279.22%	↑ 22.50%	↓ 42.82%	↑ 158.82%
YOLOv9	Download	Download	0.212	0.322	0.232	0.079	0.320	0.472	0.080	0.211	0.252	0.103	0.387	0.551	0.17
			0.187	0.295	0.199	0.130	0.242	0.171	0.079	0.231	0.310	0.234	0.396	0.239	0.44
			↓ 11.93%	↓ 8.15%	↓ 14.10%	↑ 64.98%	↓ 24.20%	↓ 63.83%	↓ 1.07%	↑ 9.85%	↑ 22.94%	↑ 127.93%	↑ 2.38%	↓ 56.64%	↑ 158.82%
YOLOv8	Download	Download	0.108	0.168	0.118	0.025	0.158	0.290	0.046	0.113	0.136	0.032	0.209	0.365	0.17
			0.121	0.193	0.130	0.070	0.163	0.149	0.056	0.158	0.211	0.131	0.281	0.220	0.47
			↑ 11.82%	↑ 14.46%	↑ 10.03%	↑ 178.14%	↑ 3.22%	↓ 48.67%	↑ 22.33%	↑ 40.05%	↑ 55.92%	↑ 308.02%	↑ 34.65%	↓ 39.72%	↑ 176.47%
YOLOv5	Download	Download	0.096	0.150	0.104	0.019	0.138	0.270	0.040	0.098	0.119	0.026	0.181	0.329	0.17
			0.109	0.174	0.116	0.059	0.150	0.134	0.050	0.139	0.188	0.111	0.254	0.205	0.54
			↑ 13.61%	↑ 16.09%	↑ 11.36%	↑ 209.43%	↑ 9.07%	↓ 50.16%	↑ 26.22%	↑ 42.71%	↑ 58.12%	↑ 329.90%	↑ 40.05%	↓ 37.62%	↑ 216.47%

Notes:

• ↑ represents percentage improvement achieved by GOIS-Det over FI-Det.
• ↓ represents performance degradation in GOIS-Det compared to FI-Det.
• Replace # in Download links with the actual URLs for downloadable files.