Encountered Accuracy Issue While Inferencing XFeat* Model with OpenVINO™ 2024.6

• Converted custom XFeat ONNX model to OpenVINO™ Intermediate Representation (IR)
• Accuracy loss while running XFeat IR model with OpenVINO™ 2024.6

XFeat.NMS function in XFeat model has a for loop that depends on the input data, and it cannot be captured correctly during export to ONNX format.

1. Upgrade OpenVINO™ version to 2025.3 and above
2. Rewrite XFeat.NMS function to avoid using the for loop for dynamic data:

   model.forward = model.detectAndCompute
      
       def NMS_no_loop(x, threshold=0.05, kernel_size=5):
           B, _, H, W = x.shape
           pad = kernel_size // 2
      
           local_max = torch.nn.functional.max_pool2d(x, kernel_size=kernel_size, stride=1, padding=pad)   
           pos = (x == local_max) & (x > threshold)   
           pos_indices = pos.nonzero(as_tuple=False)   
           batch_indices = pos_indices[:, 0]
           spatial_positions = pos_indices[:, 2:].flip(-1)   
           unique_batches, counts = torch.unique(batch_indices, return_counts=True)
           pad_val = counts.max().item()   
           pos_tensor = torch.zeros((B, pad_val, 2), dtype=torch.long, device=x.device)
      
           for b in unique_batches:
               batch_mask = batch_indices == b
               pos_tensor[b, :batch_mask.sum(), :] = spatial_positions[batch_mask]
      
           return pos_tensor
      
       model.NMS = NMS_no_loop
    
       ovm = ov.convert_model(model, example_input=(torch.randn(1,3,800,800),), verbose=True)