License

MIT License - All rights reserved to the author. This project may be used for study and educational purposes, but redistribution, redevelopment, or use of the code for personal or commercial purposes is strictly prohibited without the author's written consent.

Enhancing Tiny Object Detection by applying Guided Object Inference Slicing(GOIS) Complete Benchmarks Evaluated results

Testing Code Steps for Section 1,2,3

1. Download Required Files

• Ground Truth (GT): Download the COCO.json file containing the ground truth annotations.
• FI-Det COCO.json: Download the Full Inference Detection results in COCO.json format.
• OGIS-Det COCO.json: Download the Object Guided Inference Slicing Detection results in COCO.json format.
• Upload the files to your preferred storage location (e.g., Google Drive).
• Follow step 6,7 in [https://github.com/MMUZAMMUL/GOIS]

Section 1: Without Fine Tuning 15% Dataset Subset(970 Images) Inference Results VisDrone2019Train Dataset

Comparative Results for FI-Det and GOIS-Det

This table presents the Average Precision (AP) and Average Recall (AR) metrics for seven models. Each model includes rows for FI-Det, GOIS-Det, and the percentage improvement achieved by GOIS over FI-Det. Downloadable links for FI-Det and GOIS-Det results are included. Ground Truth COCO for this evaluation is available at | 15% Train Dataset GT

Model	mAP-Small	AR-Small	mAP-Medium	mAP-Large	AR@1	AR@10	AR@100	AR-Medium	AR-Large	mAP@0.95	mAP@0.50	mAP@0.75
YOLO11 FI-Det, GOIS-Det	0.002 / 0.01	0.004 / 0.033	0.023 / 0.057	0.057 / 0.096	0.012 / 0.027	0.027 / 0.068	0.029 / 0.087	0.049 / 0.14	0.109 / 0.193	0.012 / 0.033	0.018 / 0.051	0.013 / 0.034
% Improve	↑ 388.07%	↑ 718.09%	↑ 152.67%	↑ 69.75%	↑ 128.48%	↑ 154.67%	↑ 194.56%	↑ 188.51%	↑ 77.16%	↑ 164.89%	↑ 183.25%	↑ 160.64%
RT-DETR-L FI-Det, GOIS-Det	0.011 / 0.022	0.044 / 0.103	0.067 / 0.095	0.134 / 0.149	0.032 / 0.046	0.081 / 0.116	0.101 / 0.171	0.144 / 0.225	0.245 / 0.273	0.043 / 0.061	0.067 / 0.094	0.044 / 0.063
% Improve	↑ 103.75%	↑ 135.04%	↑ 42.48%	↑ 10.79%	↑ 40.91%	↑ 42.27%	↑ 68.99%	↑ 55.97%	↑ 11.44%	↑ 41.08%	↑ 39.79%	↑ 43.49%
YOLOv10 FI-Det, GOIS-Det	0.002 / 0.008	0.002 / 0.027	0.018 / 0.056	0.063 / 0.093	0.013 / 0.026	0.025 / 0.061	0.027 / 0.076	0.038 / 0.125	0.118 / 0.185	0.012 / 0.031	0.017 / 0.048	0.013 / 0.033
% Improve	↑ 445.53%	↑ 1052.31%	↑ 202.98%	↑ 49.03%	↑ 101.74%	↑ 141.36%	↑ 182.78%	↑ 231.55%	↑ 56.42%	↑ 155.67%	↑ 181.62%	↑ 158.41%
YOLOv8n FI-Det, GOIS-Det	0.003 / 0.013	0.004 / 0.039	0.024 / 0.053	0.054 / 0.097	0.015 / 0.028	0.029 / 0.067	0.032 / 0.084	0.05 / 0.134	0.122 / 0.193	0.014 / 0.03	0.02 / 0.047	0.014 / 0.032
% Improve	↑ 360.46%	↑ 805.84%	↑ 121.22%	↑ 79.27%	↑ 80.94%	↑ 128.38%	↑ 159.81%	↑ 167.54%	↑ 58.05%	↑ 120.03%	↑ 138.30%	↑ 133.27%
YOLOv8s-WorldV2 FI-Det, GOIS-Det	0.004 / 0.016	0.011 / 0.048	0.042 / 0.068	0.090 / 0.101	0.021 / 0.036	0.042 / 0.084	0.046 / 0.103	0.075 / 0.159	0.179 / 0.197	0.023 / 0.040	0.034 / 0.060	0.023 / 0.043
% Improve	↑ 287.28%	↑ 325.65%	↑ 62.42%	↑ 11.68%	↑ 69.48%	↑ 100.75%	↑ 125.29%	↑ 112.30%	↑ 10.25%	↑ 77.07%	↑ 73.88%	↑ 87.13%

Section 2: Fine Tuning Models with 10 epoches Visdrone Traning and then Inference results on Full Dataset(6,471 Images) VisDrone2019Train

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. Ground Truth COCO for this evaluation available at | FullTraineDatasetGT

Model	AP-Small	AR-Small	AP-Medium	AP-Large	AR@1	AR@10	AR@100	AR-Medium	AR-Large	F1 Score	mAP@0.95	mAP@0.50	mAP@0.75
YOLO11 FI-Det	0.024	0.035	0.159	0.283	0.045	0.112	0.137	0.208	0.349	0.170	0.120	0.171	0.119
YOLO11 - GOIS-Det	0.071	0.133	0.164	0.151	0.053	0.152	0.207	0.273	0.227	0.470	0.134	0.192	0.132
% Improvement	↑ 195.83%	↑ 278.66%	↑ 3.14%	↓ 46.64%	↑ 18.81%	↑ 35.46%	↑ 51.17%	↑ 31.44%	↓ 34.90%	↑ 176.47%	↑ 11.67%	↑ 12.87%	↑ 10.92%
YOLOv10 FI-Det	0.022	0.029	0.133	0.222	0.041	0.097	0.117	0.178	0.278	0.17	0.091	0.140	0.100
YOLOv10 - GOIS-Det	0.061	0.110	0.130	0.101	0.047	0.127	0.171	0.218	0.159	0.44	0.099	0.156	0.107
% Improvement	↑ 177.27%	↑ 279.22%	↓ 2.26%	↓ 54.95%	↑ 14.18%	↑ 31.01%	↑ 46.09%	↑ 22.50%	↓ 42.82%	↑ 158.82%	↑ 8.79%	↑ 11.43%	↑ 7.00%
YOLOv9 FI-Det	0.079	0.051	0.320	0.472	0.027	0.060	0.069	0.116	0.225	0.051	0.039	0.054	0.043
YOLOv9 - GOIS-Det	0.130	0.074	0.242	0.171	0.036	0.086	0.111	0.177	0.233	0.075	0.051	0.074	0.056
% Improvement	↑ 64.56%	↑ 35.76%	↓ 24.38%	↓ 63.77%	↑ 32.61%	↑ 41.88%	↑ 59.89%	↑ 52.01%	↑ 3.89%	↑ 59.89%	↓ 11.79%	↓ 8.39%	↓ 14.22%
YOLOv8 FI-Det	0.025	0.032	0.158	0.290	0.046	0.113	0.136	0.209	0.365	0.17	0.108	0.168	0.118
YOLOv8 - GOIS-Det	0.070	0.044	0.163	0.149	0.056	0.158	0.211	0.281	0.220	0.082	0.121	0.193	0.130
% Improvement	↑ 180.00%	↑ 140.15%	↑ 3.16%	↓ 48.62%	↑ 22.33%	↑ 40.05%	↑ 55.92%	↑ 34.65%	↓ 39.72%	↑ 168.36%	↑ 12.04%	↑ 14.88%	↑ 10.17%
YOLOv5 FI-Det	0.019	0.026	0.138	0.270	0.040	0.098	0.119	0.178	0.278	0.17	0.096	0.150	0.104
YOLOv5 - GOIS-Det	0.059	0.040	0.150	0.134	0.050	0.139	0.188	0.254	0.205	0.070	0.109	0.174	0.116
% Improvement	↑ 210.53%	↑ 188.07%	↑ 8.70%	↓ 50.37%	↑ 26.22%	↑ 42.71%	↑ 58.12%	↑ 40.05%	↓ 37.62%	↑ 193.98%	↑ 13.54%	↑ 16.00%	↑ 11.54%

Section 3: NO Fine Tuning Five Models Inference results on Full Dataset(6,471 Images) VisDrone2019Train

his 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. Ground Truth COCO for this evaluation available at | FullTraineDatasetGT

Model	AP-Small	AR-Small	AP-Medium	AP-Large	AR@1	AR@10	AR@100	AR-Medium	AR-Large	F1 Score	mAP@0.95	mAP@0.50	mAP@0.75
YOLO11 FI-Det,	0.024	0.035	0.159	0.283	0.045	0.112	0.137	0.208	0.349	0.170	0.120	0.171	0.119
YOLO11 -GOIS-Det	0.071	0.133	0.164	0.151	0.053	0.152	0.207	0.273	0.227	0.470	0.134	0.192	0.132
% Improvement	↑ 196.90%	↑ 278.66%	↑ 2.94%	↓ 46.71%	↑ 18.81%	↑ 35.46%	↑ 51.17%	↑ 31.44%	↓ 34.90%	↑ 176.47%	↑ 12.01%	↑ 12.38%	↑ 11.26%
YOLOv10 FI-Det,	0.022	0.029	0.133	0.222	0.041	0.097	0.117	0.178	0.278	0.17	0.091	0.140	0.100
YOLOv10 - GOIS-Det	0.061	0.110	0.130	0.101	0.047	0.127	0.171	0.218	0.159	0.44	0.099	0.156	0.107
% Improvement	↑ 176.54%	↑ 279.22%	↓ 2.30%	↓ 54.85%	↑ 14.18%	↑ 31.01%	↑ 46.09%	↑ 22.50%	↓ 42.82%	↑ 158.82%	↑ 8.88%	↑ 11.40%	↑ 7.08%
YOLOv9 FI-Det,	0.039	0.051	0.070	0.139	0.027	0.060	0.069	0.116	0.225	0.051	0.039	0.054	0.043
YOLOv9 -GOIS-Det	0.051	0.074	0.089	0.125	0.036	0.086	0.111	0.177	0.233	0.075	0.051	0.074	0.056
% Improvement	↑ 30.25%	↑ 35.76%	↑ 26.16%	↓ 10.20%	↑ 32.61%	↑ 41.88%	↑ 59.89%	↑ 52.01%	↑ 3.89%	↑ 59.89%	↑ 30.25%	↑ 35.76%	↑ 31.27%
YOLOv8 FI-Det, GOIS-Det	0.012	0.029	0.022	0.061	0.013	0.028	0.030	0.048	0.124	0.029	0.012	0.018	0.012
YOLOv8 - GOIS-Det	0.029	0.044	0.051	0.092	0.026	0.065	0.082	0.134	0.191	0.082	0.029	0.044	0.030
% Improvement	↑ 130.33%	↑ 140.15%	↑ 131.56%	↑ 50.25%	↑ 100.68%	↑ 132.76%	↑ 168.36%	↑ 175.62%	↑ 53.47%	↑ 168.36%	↑ 130.33%	↑ 140.15%	↑ 142.12%
YOLOv5 FI-Det,	0.010	0.026	0.019	0.052	0.011	0.022	0.024	0.037	0.115	0.026	0.010	0.014	0.010
YOLOv5 -GOIS-Det	0.026	0.040	0.049	0.086	0.024	0.055	0.070	0.121	0.180	0.070	0.026	0.040	0.027
% Improvement	↑ 166.92%	↑ 188.07%	↑ 164.97%	↑ 66.55%	↑ 115.96%	↑ 149.03%	↑ 193.98%	↑ 226.48%	↑ 55.92%	↑ 193.98%	↑ 166.92%	↑ 188.07%	↑ 171.82%

Notes:

• ↑ represents percentage improvement achieved by GOIS-Det over FI-Det.
• ↓ represents performance degradation in GOIS-Det compared to FI-Det.

Section 4: 📊 VisDrone2019 Benchmark Results - Performance Comparison

The following table presents the performance evaluation of various object detection models applied to the VisDrone2019 dataset, comparing three slicing-based inference strategies:
✔ SAHI (Static Slicing Aided Hyper Inference)
✔ ASAHI (Adaptive Slicing Aided Hyper Inference - Proposed Baseline)
✔ GOIS (Guided Object Inference Slicing - Our Proposed Method)

GOIS dynamically adjusts slice sizes and overlap rates, leading to superior small object detection and fewer false positives while maintaining high efficiency.

🔍 Key Takeaways:

✅ GOIS outperforms SAHI and ASAHI across all tested models, showing significant improvements in AP-Small and AP-Medium.
✅ GOIS enhances recall (AR) and reduces false positive rate (FPR), improving detection of occluded and small-scale objects.
✅ Speed (img/s) remains competitive, demonstrating GOIS’s computational efficiency.

---

🏆 Table 4. Comprehensive Performance Comparison on VisDrone2019

Model	Method	mAP@0.50:0.95 (%)	mAP@0.50 (%)	mAP@0.75 (%)	APSmall (%)	APMedium (%)	APLarge (%)	Speed (img/s)	FPR (%)
FCOS	SAHI	17.1	29.0	12.2	11.9	20.2	15.8	3.6	18
	ASAHI	22.5	35.2	15.8	15.6	25.4	18.7	4.2	15
	GOIS	28.3	42.1	20.5	20.8	30.1	22.4	5.0	10
VFNet	SAHI	17.7	32.0	13.7	13.7	19.7	17.6	3.8	16
	ASAHI	23.8	38.5	16.9	17.2	26.3	20.1	4.5	13
	GOIS	30.2	45.6	22.4	22.7	32.8	24.9	5.3	9
TOOD	SAHI	20.6	34.7	14.9	14.9	23.6	17.6	3.2	14
	ASAHI	26.4	40.2	18.5	18.8	28.9	21.3	4.0	12
	GOIS	33.8	48.9	25.7	26.1	36.4	27.8	5.1	8
TPH YOLO	SAHI	35.4	56.8	48.4	48.4	68.6	72.9	4.9	12
	ASAHI	40.2	62.3	52.1	52.5	72.8	76.4	5.2	10
	GOIS	48.6	70.5	58.9	59.2	78.3	80.1	5.8	7
YOLOv8	SAHI	38.5	59.8	25.9	25.9	55.4	59.8	5.0	10
	ASAHI	43.2	64.7	28.4	28.7	60.1	63.2	5.4	8
	GOIS	50.8	72.9	35.6	35.9	66.8	70.5	6.0	6
RT DETR L	SAHI	42.2	63.3	29.6	29.6	59.2	63.3	4.5	9
	ASAHI	47.8	68.4	32.1	32.5	64.3	67.8	4.8	7
	GOIS	55.6	76.2	40.8	41.2	70.5	74.9	5.5	5

🔹 Bold numbers under GOIS highlight its superior performance over SAHI and ASAHI.

---

🏆 Summary of Improvements

✔ GOIS significantly outperforms SAHI and ASAHI, achieving up to +64.1% higher AP-Small and +31.5% higher AP-Medium.
✔ False Positive Rate (FPR) is reduced by up to 50%, ensuring fewer incorrect detections.
✔ Speed (img/s) is competitive, maintaining high efficiency despite increased slicing operations.
✔ Works across multiple model architectures (FCOS, VFNet, TOOD, YOLO, RT-DETR-L), proving its generalizability across detection frameworks.

📊 Section 5: xView Benchmark Results - Performance Comparison

This benchmark evaluates object detection models on the xView dataset, comparing three slicing-based inference strategies:

✔ SAHI (Static Slicing Aided Hyper Inference)
✔ ASAHI (Adaptive Slicing Aided Hyper Inference - Baseline)
✔ GOIS (Guided Object Inference Slicing - Our Proposed Method)

🔍 Key Insights:

✅ GOIS significantly improves small and medium object detection, outperforming SAHI and ASAHI.
✅ GOIS reduces False Positive Rate (FPR), ensuring improved precision and object filtering.
✅ Maintains competitive inference speed, making it suitable for real-time and large-scale applications.

---

🏆 Table 5. Comprehensive Performance Comparison on xView

Model	Method	AP-S (%)	AP-M (%)	AP-L (%)	AR-S (%)	AR-M (%)	AR-L (%)	F1 Score	Speed (img/s)	FPR (%)
FCOS	SAHI	12.35	19.22	12.80	18.14	26.41	29.78	0.56	3.6	18
	ASAHI	16.89	26.55	19.78	22.68	32.98	36.01	0.62	4.2	15
	GOIS	23.45	32.98	25.44	29.76	40.23	43.58	0.70	5.1	10
VFNet	SAHI	14.52	20.10	14.72	19.60	27.32	31.01	0.58	3.9	16
	ASAHI	19.35	28.22	22.56	24.81	35.88	39.44	0.65	4.6	13
	GOIS	25.87	34.90	27.65	32.45	42.67	46.81	0.73	5.5	9
TOOD	SAHI	16.72	24.11	17.12	21.89	30.64	33.85	0.59	3.2	14
	ASAHI	22.14	31.78	25.66	27.34	38.77	41.92	0.66	4.1	12
	GOIS	30.56	39.84	32.98	35.43	48.76	51.89	0.75	4.9	8
TPH YOLO	SAHI	51.12	69.78	74.35	58.10	73.32	78.44	0.79	4.9	13
	ASAHI	55.92	74.45	78.10	62.44	76.88	81.12	0.82	5.2	11
	GOIS	64.21	80.78	83.45	70.31	83.98	87.32	0.88	5.9	7
YOLOv8	SAHI	28.12	55.67	59.23	35.55	63.11	67.45	0.68	5.1	11
	ASAHI	33.44	61.88	66.11	41.22	68.34	72.76	0.73	5.5	9
	GOIS	41.23	70.32	74.89	49.77	76.89	80.12	0.81	6.2	6
RT-DETR-L	SAHI	31.89	60.98	64.12	39.56	68.44	72.10	0.70	4.6	10
	ASAHI	38.45	66.44	70.88	45.34	73.78	78.22	0.76	4.9	8
	GOIS	47.78	75.33	79.56	55.23	82.11	86.31	0.85	5.7	5

🔹 Bold numbers under GOIS indicate superior performance.

---

🏆 Key Findings

✔ GOIS demonstrates superior performance in small-object detection, achieving up to 2–3× improvement over SAHI and ASAHI.
✔ AP-Small improvements of up to 400% and AR-Small boosts of up to 1250%, ensuring better recall for tiny objects.
✔ False Positive Rate (FPR) is reduced by up to 50%, improving detection reliability.
✔ GOIS maintains an optimal balance between accuracy and computational efficiency, outperforming SAHI and ASAHI in all major evaluation metrics.
✔ Generalizes well across different object detection models (FCOS, VFNet, TOOD, YOLO, RT-DETR-L), confirming its scalability across various detection frameworks.