🚗 Road Object Detection in Fish-Eye Cameras

Hung quan | Van Hoang | My Kim | Thien Nhan | Thanh Dat

---

iccv.webm

Inference result samples on Fisheye1K using 640x640_fisheye8k.engine (FP32).

---

👾 System Information

• Platform: Jetson AGX Xavier (JetPack 5.1.2, L4T R35.4.1)
• TensorRT Version: 8.5.0.2
• Torch Version: 2.1.0 (torch-2.1.0a0+41361538.nv23.06-cp38-cp38-linux_aarch64)
• Torchvision Version: v0.16.1
• Input Resolution: 1024×1024 or 640×640

🔧 You can adjust parameters via the configuration file: config/config.yaml.

---

📊 Evaluation Metrics (NVIDIA AI Challenge 2025)

Model	AP0.5:0.95	AP0.5	APS	APM	APL	F1 Score
1024×1024_fisheye8k + 1024×1024_visdra_m (best)	0.5238	0.7226	0.3369	0.6877	0.5925	0.6139
640×640_fisheye8k (FP32)	0.5556	0.7915	0.3810	0.6880	0.5727	0.5995

---

⚡ Inference Speed on Jetson AGX Xavier (30W ALL, single .engine)

Accuracy for the FP32 model is reported based on the ICCV 2025 evaluation. The FP16 model has not yet been evaluated for accuracy; results shown here reflect FPS only.

Model	FPS	Normalized (max=25)
640×640_fisheye8k (FP32)	12.09	0.4836
640×640_fisheye8k (FP16)	21.89	0.8756

---

📥 Pretrained Weights

FP16-trained weights are currently not available.

Model (FP32)	640×640 Weights	1024×1024 Weights
dfine_hgnetv2_m_fisheye8k	Download	Download
dfine_hgnetv2_m_visdra	Download	Download

---

🛠️ TensorRT Engine Build

We use the following command to build .engine files:

⚠️ Note: Using --fp16 or --int8 on FP32-trained models may cause numerical overflow.

trtexec \
  --onnx=model/dfine_640.onnx \
  --saveEngine=model/dfine_640.engine \
  --memPoolSize=workspace:11000 \
  --useCudaGraph \
  --best \
  --minShapes=images:1x3x640x640,orig_target_sizes:1x2 \
  --optShapes=images:1x3x640x640,orig_target_sizes:1x2 \
  --maxShapes=images:1x3x640x640,orig_target_sizes:1x2

---

❤️ Acknowledgement

This work is built upon the amazing D-FINE project.