Add support for exporting to ONNX.

[acai66]

1. Support for exporting xfeat, xfeat+matching models.
2. Support for exporting dynamic shapes.
3. Support for exporting a specified version of opset.
4. Add onnxruntime inference demo.

Model	Inputs	Outputs	Note
xfeat.onnx	images	feats, keypoints, heatmaps	Extract image keypoints and features
xfeat_dualscale.onnx	images	mkpts, feats, sc	Extract dualscale image keypoints and features.
matching.onnx	mkpts0, feats0, sc0, mkpts1, feats1	matches, batch_indexes	Match dualscale keypoints and features
xfeat_matching.onnx	images0, images1	matches, batch_indexes	End-to-end extraction of features from two sets of images and performing feature matching.

onnx_models.zip

examples

• FPS is limited by this line, especially when a large number of keypoints are matched https://github.com/verlab/accelerated_features/blob/b2f6b99b85db0672fc5e49df016850370ce8ad57/realtime_demo.py#L239

xfeat_onnxruntime.py

import numpy as np
import onnxruntime as ort


def create_ort_session(model_path, trt_engine_cache_path='trt_engine_cache', trt_engine_cache_prefix='model'):
    tmp_ort_session = ort.InferenceSession(model_path, providers=['CPUExecutionProvider'])

    # print the input,output names and shapes
    for i in range(len(tmp_ort_session.get_inputs())):
        print(f"Input name: {tmp_ort_session.get_inputs()[i].name}, shape: {tmp_ort_session.get_inputs()[i].shape}")
    for i in range(len(tmp_ort_session.get_outputs())):
        print(f"Output name: {tmp_ort_session.get_outputs()[i].name}, shape: {tmp_ort_session.get_outputs()[i].shape}")

    providers = [
        # The TensorrtExecutionProvider is the fastest.
        ('TensorrtExecutionProvider', { 
            'device_id': 0,
            'trt_max_workspace_size': 4 * 1024 * 1024 * 1024,
            'trt_fp16_enable': True,
            'trt_engine_cache_enable': True,
            'trt_engine_cache_path': trt_engine_cache_path,
            'trt_engine_cache_prefix': trt_engine_cache_prefix,
            'trt_dump_subgraphs': False,
            'trt_timing_cache_enable': True,
            'trt_timing_cache_path': trt_engine_cache_path,
            #'trt_builder_optimization_level': 3,
        }),

        # The CUDAExecutionProvider is slower than PyTorch, 
        # possibly due to performance issues with large matrix multiplication "cossim = torch.bmm(feats1, feats2.permute(0,2,1))"
        # Reducing the top_k value when exporting to ONNX can decrease the matrix size.
        ('CUDAExecutionProvider', { 
            'device_id': 0,
            'gpu_mem_limit': 4 * 1024 * 1024 * 1024,
        }),
        ('CPUExecutionProvider',{ 
        })
    ]
    ort_session = ort.InferenceSession(model_path, providers=providers)

    return ort_session


class XFeat:

    def __init__(self, xfeat_model_path='./xfeat_dualscale.onnx', matcher_model_path='./matching.onnx'):
        self.xfeat_ort_session = create_ort_session(xfeat_model_path, trt_engine_cache_prefix='xfeat_dualscale')
        self.matcher_ort_session = create_ort_session(matcher_model_path, trt_engine_cache_prefix='matching')

        # warm up
        for i in range(5):
            image = np.zeros((640, 640, 3), dtype=np.float32)
            self.detectAndCompute(image)
            mkpts0 = np.zeros((1, 4800, 2), dtype=np.float32)
            feats0 = np.zeros((1, 4800, 64), dtype=np.float32)
            sc0 = np.zeros((1, 4800), dtype=np.float32)
            mkpts1 = np.zeros((1, 4800, 2), dtype=np.float32)
            feats1 = np.zeros((1, 4800, 64), dtype=np.float32)
            self.match(mkpts0, feats0, sc0, mkpts1, feats1)

    def detectAndCompute(self, image_data, mask=None):
        input_array = np.expand_dims(image_data.transpose((2, 0, 1)) , axis=0).astype(np.float32)
        inputs = {
            self.xfeat_ort_session.get_inputs()[0].name: input_array,
        }
        mkpts0, feats0, sc = self.xfeat_ort_session.run(None, inputs)

        return {
            "keypoints": mkpts0,
            "descriptors": feats0,
            "sc": sc,
        }

    def match(self, mkpts0, feats0, sc0, mkpts1, feats1):
        inputs = {
            self.matcher_ort_session.get_inputs()[0].name: mkpts0,
            self.matcher_ort_session.get_inputs()[1].name: feats0,
            self.matcher_ort_session.get_inputs()[2].name: sc0,
            self.matcher_ort_session.get_inputs()[3].name: mkpts1,
            self.matcher_ort_session.get_inputs()[4].name: feats1,
        }
        matches, batch_indexes = self.matcher_ort_session.run(None, inputs)

        return matches, batch_indexes

realtime_demo.py patches

diff --git a/realtime_demo.py b/realtime_demo.py
index 6c867fd..8e21ae2 100644
--- a/realtime_demo.py
+++ b/realtime_demo.py
@@ -7,20 +7,18 @@
 
 import cv2
 import numpy as np
-import torch

 from time import time, sleep
 import argparse, sys, tqdm
 import threading

-from modules.xfeat import XFeat

 def argparser():
     parser = argparse.ArgumentParser(description="Configurations for the real-time matching demo.")
     parser.add_argument('--width', type=int, default=640, help='Width of the video capture stream.')
     parser.add_argument('--height', type=int, default=480, help='Height of the video capture stream.')
     parser.add_argument('--max_kpts', type=int, default=3_000, help='Maximum number of keypoints.')
-    parser.add_argument('--method', type=str, choices=['ORB', 'SIFT', 'XFeat'], default='XFeat', help='Local feature detection method to use.')
+    parser.add_argument('--method', type=str, choices=['ORB', 'SIFT', 'XFeat', 'XFeat_Ort'], default='XFeat', help='Local feature detection method to use.')
     parser.add_argument('--cam', type=int, default=0, help='Webcam device number.')
     return parser.parse_args()

@@ -52,6 +50,7 @@ class CVWrapper():
     def __init__(self, mtd):
         self.mtd = mtd
     def detectAndCompute(self, x, mask=None):
+        import torch
         return self.mtd.detectAndCompute(torch.tensor(x).permute(2,0,1).float()[None])[0]

 class Method:
@@ -65,7 +64,12 @@ def init_method(method, max_kpts):
     elif method == "SIFT":
         return Method(descriptor=cv2.SIFT_create(max_kpts, contrastThreshold=-1, edgeThreshold=1000), matcher=cv2.BFMatcher(cv2.NORM_L2, crossCheck=True))
     elif method == "XFeat":
+        from modules.xfeat import XFeat
         return Method(descriptor=CVWrapper(XFeat(top_k = max_kpts)), matcher=XFeat())
+    elif method == "XFeat_Ort":
+        from xfeat_onnxruntime import XFeat
+        xfeat = XFeat()
+        return Method(descriptor=xfeat, matcher=xfeat)
     else:
         raise RuntimeError("Invalid Method.")

@@ -200,6 +204,12 @@ class MatchingDemo:
         if self.args.method in ['SIFT', 'ORB']:
             kp1, des1 = self.ref_precomp
             kp2, des2 = self.method.descriptor.detectAndCompute(current_frame, None)
+        elif self.args.method in ['XFeat_Ort']:
+            current = self.method.descriptor.detectAndCompute(current_frame)
+            kpts1, descs1, sc1 = self.ref_precomp['keypoints'], self.ref_precomp['descriptors'], self.ref_precomp['sc']
+            kpts2, descs2 = current['keypoints'], current['descriptors']
+            matches, batch_indexes = self.method.matcher.match(kpts1, descs1, sc1, kpts2, descs2)
+            points1, points2 = matches[batch_indexes == 0][..., :2], matches[batch_indexes == 0][..., 2:]
         else:
             current = self.method.descriptor.detectAndCompute(current_frame)
             kpts1, descs1 = self.ref_precomp['keypoints'], self.ref_precomp['descriptors']

[IamShubhamGupto]

Im slightly busy with graduation this week, I will be back to work on this by the weekend

[acai66]

Im slightly busy with graduation this week, I will be back to work on this by the weekend

No problem, thanks for letting me know. Congratulations on your graduation!

[acai66]

Here is a sample python code for running xfeat_matching.onnx with TensorRT API. This example does not include exception handling, input/output validation, etc., and is for reference only.

xfeat_tensorrt_python.zip

update 2024-06-24:

• Update xfeat_dualscale.onnx, Add support for fewer than 4800 num_keypoints to accommodate low-resolution images
• Update matching.onnx, Add support for different numbers of two sets of features
• Add tensorrt demo codes for xfeat_dualscale.onnx and matching.onnx

import os
import sys
import time
from typing import Optional, List
from functools import reduce

import numpy as np
import cv2
import tensorrt as trt
import cupy as cp    # pip install cupy-cuda12x
from cuda import cuda, cudart    # pip install cuda-python
from packaging.version import Version


if Version(trt.__version__) >= Version('9.0.0'):
    # This is a simple ASCII-art progress monitor comparable to the C++ version in sample_progress_monitor.
    class SimpleProgressMonitor(trt.IProgressMonitor):
        def __init__(self):
            trt.IProgressMonitor.__init__(self)
            self._active_phases = {}
            self._step_result = True

        def phase_start(self, phase_name, parent_phase, num_steps):
            try:
                if parent_phase is not None:
                    nbIndents = 1 + self._active_phases[parent_phase]['nbIndents']
                else:
                    nbIndents = 0
                self._active_phases[phase_name] = { 'title': phase_name, 'steps': 0, 'num_steps': num_steps, 'nbIndents': nbIndents }
                self._redraw()
            except KeyboardInterrupt:
                # The phase_start callback cannot directly cancel the build, so request the cancellation from within step_complete.
                _step_result = False

        def phase_finish(self, phase_name):
            try:
                del self._active_phases[phase_name]
                self._redraw(blank_lines=1) # Clear the removed phase.
            except KeyboardInterrupt:
                _step_result = False

        def step_complete(self, phase_name, step):
            try:
                self._active_phases[phase_name]['steps'] = step
                self._redraw()
                return self._step_result
            except KeyboardInterrupt:
                # There is no need to propagate this exception to TensorRT. We can simply cancel the build.
                return False

        def _redraw(self, *, blank_lines=0):
            # The Python curses module is not widely available on Windows platforms.
            # Instead, this function uses raw terminal escape sequences. See the sample documentation for references.
            def clear_line():
                print('\x1B[2K', end='')
            def move_to_start_of_line():
                print('\x1B[0G', end='')
            def move_cursor_up(lines):
                print('\x1B[{}A'.format(lines), end='')

            def progress_bar(steps, num_steps):
                INNER_WIDTH = 10
                completed_bar_chars = int(INNER_WIDTH * steps / float(num_steps))
                return '[{}{}]'.format(
                    '=' * completed_bar_chars,
                    '-' * (INNER_WIDTH - completed_bar_chars))

            # Set max_cols to a default of 200 if not run in interactive mode.
            max_cols = os.get_terminal_size().columns if sys.stdout.isatty() else 200

            move_to_start_of_line()
            for phase in self._active_phases.values():
                phase_prefix = '{indent}{bar} {title}'.format(
                    indent = ' ' * phase['nbIndents'],
                    bar = progress_bar(phase['steps'], phase['num_steps']),
                    title = phase['title'])
                phase_suffix = '{steps}/{num_steps}'.format(**phase)
                allowable_prefix_chars = max_cols - len(phase_suffix) - 2
                if allowable_prefix_chars < len(phase_prefix):
                    phase_prefix = phase_prefix[0:allowable_prefix_chars-3] + '...'
                clear_line()
                print(phase_prefix, phase_suffix)
            for line in range(blank_lines):
                clear_line()
                print()
            move_cursor_up(len(self._active_phases) + blank_lines)
            sys.stdout.flush()

FP16_ENABLE = True
EXPLICIT_BATCH = 1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)

def check_cuda_err(err):
    if isinstance(err, cuda.CUresult):
        if err != cuda.CUresult.CUDA_SUCCESS:
            raise RuntimeError("Cuda Error: {}".format(err))
    if isinstance(err, cudart.cudaError_t):
        if err != cudart.cudaError_t.cudaSuccess:
            raise RuntimeError("Cuda Runtime Error: {}".format(err))
    else:
        raise RuntimeError("Unknown error type: {}".format(err))


def cuda_call(call):
    err, res = call[0], call[1:]
    check_cuda_err(err)
    if len(res) == 1:
        res = res[0]
    return res


class OutputAllocator(trt.IOutputAllocator):
    def __init__(self):
        trt.IOutputAllocator.__init__(self)
        self.buffers = {}
        self.shapes = {}

    def reallocate_output(self, tensor_name, memory, size, alignment):
        output_dtype = cp.dtype(cp.byte)
        output = cp.empty(size, output_dtype)
        ptr = output.data.ptr
        self.buffers[tensor_name] = output
        return ptr

    def notify_shape(self, tensor_name, shape):
        self.shapes[tensor_name] = tuple(shape)


def get_engine(onnx_file_path, engine_file_path=""):
    """Attempts to load a serialized engine if available, otherwise builds a new TensorRT engine and saves it."""

    TRT_LOGGER = trt.Logger()

    def build_engine():
        """Takes an ONNX file and creates a TensorRT engine to run inference with"""
        with trt.Builder(TRT_LOGGER) as builder, builder.create_network(
            0 if Version(trt.__version__) >= Version('9.0.0') else EXPLICIT_BATCH
        ) as network, builder.create_builder_config() as config, trt.OnnxParser(
            network, TRT_LOGGER
        ) as parser, trt.Runtime(
            TRT_LOGGER
        ) as runtime:
            # Parse model file
            if not os.path.exists(onnx_file_path):
                print(
                    "ONNX file {} not found.".format(onnx_file_path)
                )
                return None
            print("Loading ONNX file from path {}...".format(onnx_file_path))
            with open(onnx_file_path, "rb") as model:
                print("Beginning ONNX file parsing")
                if not parser.parse(model.read()):
                    print("ERROR: Failed to parse the ONNX file.")
                    for error in range(parser.num_errors):
                        print(parser.get_error(error))
                    return None
            
            print("Completed parsing of ONNX file")

            if Version(trt.__version__) >= Version('9.0.0'):
                config.progress_monitor = SimpleProgressMonitor()
            config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, 4 << 30) # 4GB
            # config.max_workspace_size = 1 << 30  # 1GB tensorrt < 8.4
            # builder.max_batch_size = 1
            
            for i in range(network.num_inputs):
                input_node = network.get_input(i)
                print('input ', i, input_node.name, input_node.shape)
            assert network.num_inputs == 2, "For xfeat_matching.onnx, only supports two input nodes."

            profile = builder.create_optimization_profile()
            profile.set_shape(network.get_input(0).name, (1, 3, 64, 64), (8, 3, 640, 640), (32, 3, 1280, 1280))
            profile.set_shape(network.get_input(1).name, (1, 3, 64, 64), (8, 3, 640, 640), (32, 3, 1280, 1280))
            config.add_optimization_profile(profile)
            
            if builder.platform_has_fast_fp16 and FP16_ENABLE:
                print("FP16 mode enabled")
                config.set_flag(trt.BuilderFlag.FP16)
            
            print("Building an engine from file {}; this may take a while...".format(onnx_file_path))
            plan = builder.build_serialized_network(network, config)
            engine = runtime.deserialize_cuda_engine(plan)
            print("Completed creating Engine")
            with open(engine_file_path, "wb") as f:
                f.write(plan)
            return engine

    if os.path.exists(engine_file_path):
        # If a serialized engine exists, use it instead of building an engine.
        print("Reading engine from file {}".format(engine_file_path))
        try:
            with open(engine_file_path, "rb") as f, trt.Runtime(TRT_LOGGER) as runtime:
                engine = runtime.deserialize_cuda_engine(f.read())
                if engine == None:
                    print("Deserialization of the engine from {} failed. Falling back to building the engine".format(engine_file_path))
                    return build_engine()
                return engine
        except Exception as e:
            print(e)
            print("Deserialization of the engine from {} failed. Falling back to building the engine".format(engine_file_path))
            return build_engine()
    else:
        return build_engine()


# This function is generalized for multiple inputs/outputs for full dimension networks.
# inputs and outputs are expected to be lists of HostDeviceMem objects.
def do_inference(context, engine, inputs_dict, output_allocator, stream):
    num_io = engine.num_io_tensors
    outputs = []
    for i in range(num_io):
        name = engine.get_tensor_name(i)
        if engine.get_tensor_mode(name) == trt.TensorIOMode.INPUT:
            context.set_tensor_address(name, inputs_dict[name].data.ptr)
            context.set_input_shape(name, inputs_dict[name].shape)
    context.execute_async_v3(stream_handle=stream)
    cuda_call(cudart.cudaStreamSynchronize(stream))

    for i in range(num_io):
        name = engine.get_tensor_name(i)
        if engine.get_tensor_mode(name) == trt.TensorIOMode.OUTPUT:
            output_shape = output_allocator.shapes[name]
            output_dtype = cp.dtype(trt.nptype(engine.get_tensor_dtype(name)))
            output_size = output_dtype.itemsize * reduce((lambda x, y: x * y), output_shape, 1)
            outputs.append(output_allocator.buffers[name][:output_size].view(output_dtype).reshape(output_shape))

    return outputs


class XFEAT:

    def __init__(self, modelPath, gpu_id=0) -> None:

        self.gpu_id = gpu_id

        # Init use gpu_id
        cudart.cudaSetDevice(self.gpu_id)
        
        engine_file_path = os.path.splitext(modelPath)[0] + ".trt"
        self.engine = get_engine(modelPath, engine_file_path)
        self.context = self.engine.create_execution_context()
        self.stream = cuda_call(cudart.cudaStreamCreate())

        self.output_allocator = OutputAllocator()
        for i in range(self.engine.num_io_tensors):
            name = self.engine.get_tensor_name(i)
            if self.engine.get_tensor_mode(name) == trt.TensorIOMode.OUTPUT:
                self.context.set_output_allocator(name, self.output_allocator)

        tensor_names = [self.engine.get_tensor_name(i) for i in range(self.engine.num_io_tensors)]
        for tensor_name in tensor_names:
            print(tensor_name, self.engine.get_tensor_shape(tensor_name))
    

    def inference_preprocessed(self, input_array_0, input_array_1):
        cudart.cudaSetDevice(self.gpu_id)

        inputs_dict = {
            "images0": input_array_0,
            "images1": input_array_1,
        }

        trt_outputs = do_inference(self.context, self.engine, inputs_dict=inputs_dict, output_allocator=self.output_allocator, stream=self.stream)
        matches = cp.asnumpy(trt_outputs[0])
        batch_indexes = cp.asnumpy(trt_outputs[1])
        
        return matches, batch_indexes

        
    def __call__(self, input_array_0, input_array_1):
        input_array_device_0 = cp.asarray(input_array_0, dtype=cp.float32)
        input_array_device_1 = cp.asarray(input_array_1, dtype=cp.float32)

        return self.inference_preprocessed(input_array_device_0, input_array_device_1)
    

    def __del__(self) -> None:
        if hasattr(self, 'stream'):
            if cudart != None:
                cuda_call(cudart.cudaStreamDestroy(self.stream))


def warp_corners_and_draw_matches(ref_points, dst_points, img1, img2):
    # Calculate the Homography matrix
    H, mask = cv2.findHomography(ref_points, dst_points, cv2.USAC_MAGSAC, 3.5, maxIters=1_000, confidence=0.999)
    mask = mask.flatten()

    # Get corners of the first image (image1)
    h, w = img1.shape[:2]
    corners_img1 = np.array([[0, 0], [w-1, 0], [w-1, h-1], [0, h-1]], dtype=np.float32).reshape(-1, 1, 2)

    # Warp corners to the second image (image2) space
    warped_corners = cv2.perspectiveTransform(corners_img1, H)

    # Draw the warped corners in image2
    img2_with_corners = img2.copy()
    for i in range(len(warped_corners)):
        start_point = tuple(warped_corners[i-1][0].astype(int))
        end_point = tuple(warped_corners[i][0].astype(int))
        cv2.line(img2_with_corners, start_point, end_point, (0, 255, 0), 4)  # Using solid green for corners

    # Prepare keypoints and matches for drawMatches function
    keypoints1 = [cv2.KeyPoint(p[0], p[1], 5) for p in ref_points]
    keypoints2 = [cv2.KeyPoint(p[0], p[1], 5) for p in dst_points]
    matches = [cv2.DMatch(i,i,0) for i in range(len(mask)) if mask[i]]

    # Draw inlier matches
    img_matches = cv2.drawMatches(img1, keypoints1, img2_with_corners, keypoints2, matches, None,
                                  matchColor=(0, 255, 0), flags=2)

    return img_matches


if __name__ == "__main__":
    onnx_file_path = './models/xfeat_matching.onnx'
    input_image_0_path = './images/ref.png'
    input_image_1_path = './images/tgt.png'

    xfeat = XFEAT(onnx_file_path, 0)
    image_0 = cv2.imread(input_image_0_path, cv2.IMREAD_COLOR)
    image_1 = cv2.imread(input_image_1_path, cv2.IMREAD_COLOR)

    batch_zise = 8  # Psuedo-batch the input images
    input_array_0 = np.expand_dims(image_0.transpose(2, 0, 1), axis=0).repeat(batch_zise, axis=0)
    input_array_1 = np.expand_dims(image_1.transpose(2, 0, 1), axis=0).repeat(batch_zise, axis=0)

    matches, batch_indexes = xfeat(input_array_0, input_array_1)
    mkpts_0, mkpts_1 = matches[batch_indexes == 0][..., :2], matches[batch_indexes == 0][..., 2:]

    img_matches = warp_corners_and_draw_matches(mkpts_0, mkpts_1, image_0, image_1)
    cv2.imshow("Matches", img_matches)
    cv2.waitKey(0)
    

    loop = 100
    start = time.time()
    for i in range(loop):
        matches, batch_indexes = xfeat(input_array_0, input_array_1)
    end = time.time()
    print("Time: ", (end - start) / loop / batch_zise)
    print("FPS: ", batch_zise * loop / (end - start))

[guipotje]

Hi @acai66 @IamShubhamGupto ,

Thank you very much for providing the ONNX examples in both C++ and Python. They look amazing and incredibly useful!

Are you still waiting for @IamShubhamGupto to review the merge? I am not sure if you both merged the work you did last month, so I am just checking in to see if I should start reviewing the PR.

Once again, thank you @acai66 and @IamShubhamGupto for the ONNX examples. I appreciate your effort to make XFeat deployment much better!

[acai66]

Hi @acai66 @IamShubhamGupto ,

Thank you very much for providing the ONNX examples in both C++ and Python. They look amazing and incredibly useful!

Are you still waiting for @IamShubhamGupto to review the merge? I am not sure if you both merged the work you did last month, so I am just checking in to see if I should start reviewing the PR.

Once again, thank you @acai66 and @IamShubhamGupto for the ONNX examples. I appreciate your effort to make XFeat deployment much better!

Thank you for checking in. I am still waiting for @IamShubhamGupto's review. Unfortunately, I haven't received any response from @IamShubhamGupto regarding last month's work, which might be due to the busy graduation season.

[IamShubhamGupto]

Hi @acai66 @IamShubhamGupto , Thank you very much for providing the ONNX examples in both C++ and Python. They look amazing and incredibly useful! Are you still waiting for @IamShubhamGupto to review the merge? I am not sure if you both merged the work you did last month, so I am just checking in to see if I should start reviewing the PR. Once again, thank you @acai66 and @IamShubhamGupto for the ONNX examples. I appreciate your effort to make XFeat deployment much better!

Thank you for checking in. I am still waiting for @IamShubhamGupto's review. Unfortunately, I haven't received any response from @IamShubhamGupto regarding last month's work, which might be due to the busy graduation season.

Hey @acai66 @guipotje sorry to keep both of you waiting. I did just graduate and was busy with a few conferences and competitions. As for the development on this branch, I believe you should go ahead and merge this branch. I will be back on contributing to this project some time later.

[noahzn]

@acai66 Hi, thank you for your contribution. I found that only detectAndComputeDense is supported. Could you also make it support for detectAndCompute?

[acai66]

@acai66 Hi, thank you for your contribution. I found that only detectAndComputeDense is supported. Could you also make it support for detectAndCompute?

commit: https://github.com/verlab/accelerated_features/pull/5/commits/f4a55c14fddbf61ea1f6de83c8a30baacecdc88b

[noahzn]

@acai66 Thank you very much!

[noahzn]

@acai66 Have you met a problem in DetectAndCompute that the #Select top-k features function exceeded the max number when using TRT Execution Provider.

[stschake]

Hey @acai66,

thanks for all your work on the ONNX export of XFeat, it's been very handy. @guipotje recently added the Lighterglue addon matcher so I spent some time to make that available for ONNX export on top of your changes, see the branch here:

https://github.com/stschake/accelerated_features/tree/feature/lighterglue-onnx

The upstream code uses kornia which isn't suitable for ONNX export, so I started with the LightGlue-ONNX implementation and modified it slightly to add things like keypoints normalization directly in the model, in the xfeat tradition.

[yrik]

Thanks guys! Looking forward for onnx version.