rdk_model_zoo/
|-- samples/
|   |-- robotics/
|   |   `-- himloco/              # Unitree Go2 locomotion policy
|   |-- vision/
|   |   |-- clip/                 # Image-text multimodal matching
|   |   |-- convnext/             # Image classification
|   |   |-- edgenext/             # Image classification
|   |   |-- efficientformer/      # Image classification
|   |   |-- efficientformerv2/    # Image classification
|   |   |-- efficient_sam/        # Promptable image segmentation
|   |   |-- mobile_sam/           # Promptable image segmentation
|   |   |-- efficientnet/         # Image classification
|   |   |-- efficientvit/         # Image classification
|   |   |-- fasternet/            # Image classification
|   |   |-- fastvit/              # Image classification
|   |   |-- fcos/                 # Object detection
|   |   |-- googlenet/            # Image classification
|   |   |-- hgnetv2/              # Image classification
|   |   |-- lprnet/               # License plate recognition
|   |   |-- mobilenetv1/          # Image classification
|   |   |-- mobilenetv2/          # Image classification
|   |   |-- mobilenetv3/          # Image classification
|   |   |-- mobilenetv4/          # Image classification
|   |   |-- mobileone/            # Image classification
|   |   |-- modnet/               # Image matting
|   |   |-- paddleocr/            # OCR text detection and recognition
|   |   |-- pp_liteseg/           # Semantic segmentation
|   |   |-- repghost/             # Image classification
|   |   |-- repvgg/               # Image classification
|   |   |-- repvit/               # Image classification
|   |   |-- resnet/               # Image classification
|   |   |-- resnext/              # Image classification
|   |   |-- unet/                 # Semantic segmentation
|   |   |-- ultralytics_yolo/     # Detection, segmentation, pose, classification
|   |   |-- ultralytics_yolo26/   # Detection, segmentation, pose, classification
|   |   |-- vargconvnet/          # Image classification
|   |   |-- yolo26_depth/         # Monocular depth estimation
|   |   |-- yolov5/               # Object detection
|   |   `-- yoloworld/           # Open-vocabulary object detection
|-- docs/                  # Project guidelines and reference documentation
|   `-- tros/              # TROS integration guides and examples
|-- datasets/              # Sample datasets and download scripts
|-- skills/                # RDK Model Zoo skills (placeholder)
|-- utils/                 # Shared C++ / Python utilities

• Ensure OpenExplorer Docker and board-side libdnn.so versions are up-to-date.
• Check if model export followed the structure adjustments/operator replacements required in the model's README.
• Verify cosine similarity of each output node is >= 0.999 (minimum 0.99) during quantization validation.

• Python API performance is lower than C/C++. For maximum performance, use C/C++.
• Benchmark data (pure forward) excludes pre/post-processing. Models with NV12 input usually achieve peak BPU throughput.
• Ensure CPU/BPU frequency is locked to maximum.
• Check for other resource-heavy processes.

• Refer to the PTQ accuracy debugging section in the platform documentation.
• If INT8 loss is severe due to model characteristics, consider Mixed Precision or QAT (Quantization-Aware Training).

Update the resolution in preprocess.py to match your ONNX model's input size. Delete old calibration data and re-run the calibration script.

• Deployment uses fixed shape and INT8 quantization, unlike dynamic shape/float official tests.
• Slight implementation differences in evaluation scripts (e.g., pycocotools).
• NCHW-RGB to NV12 conversion adds minimal pixel-level loss.

Yes. Non-quantizable or BPU-unsupported operators fallback to CPU. Even for pure BPU models, input/output quantization/dequantization nodes are executed by the CPU.