first commit

tests/models/vitpose/test_image_processing_vitpose.py

@@ -0,0 +1,293 @@
+# Copyright 2024 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import io
+import unittest
+
+import httpx
+import numpy as np
+import pytest
+
+from transformers.testing_utils import require_torch, require_torch_accelerator, require_vision
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
+
+
+if is_torch_available():
+    import torch
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+class VitPoseImageProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=400,
+        do_affine_transform=True,
+        size=None,
+        do_rescale=True,
+        rescale_factor=1 / 255,
+        do_normalize=True,
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+    ):
+        size = size if size is not None else {"height": 20, "width": 20}
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.do_affine_transform = do_affine_transform
+        self.size = size
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+
+    def prepare_image_processor_dict(self):
+        return {
+            "do_affine_transform": self.do_affine_transform,
+            "size": self.size,
+            "do_rescale": self.do_rescale,
+            "rescale_factor": self.rescale_factor,
+            "do_normalize": self.do_normalize,
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+        }
+
+    def expected_output_image_shape(self, images):
+        return self.num_channels, self.size["height"], self.size["width"]
+
+    def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
+        return prepare_image_inputs(
+            batch_size=self.batch_size,
+            num_channels=self.num_channels,
+            min_resolution=self.min_resolution,
+            max_resolution=self.max_resolution,
+            equal_resolution=equal_resolution,
+            numpify=numpify,
+            torchify=torchify,
+        )
+
+
+@require_torch
+@require_vision
+class VitPoseImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = VitPoseImageProcessingTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    def test_image_processor_properties(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processing = image_processing_class(**self.image_processor_dict)
+            self.assertTrue(hasattr(image_processing, "do_affine_transform"))
+            self.assertTrue(hasattr(image_processing, "size"))
+            self.assertTrue(hasattr(image_processing, "do_rescale"))
+            self.assertTrue(hasattr(image_processing, "rescale_factor"))
+            self.assertTrue(hasattr(image_processing, "do_normalize"))
+            self.assertTrue(hasattr(image_processing, "image_mean"))
+            self.assertTrue(hasattr(image_processing, "image_std"))
+
+    def test_image_processor_from_dict_with_kwargs(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processor = image_processing_class.from_dict(self.image_processor_dict)
+            self.assertEqual(image_processor.size, {"height": 20, "width": 20})
+
+            image_processor = image_processing_class.from_dict(
+                self.image_processor_dict, size={"height": 42, "width": 42}
+            )
+            self.assertEqual(image_processor.size, {"height": 42, "width": 42})
+
+    def test_call_pil(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random PIL images
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False)
+            for image in image_inputs:
+                self.assertIsInstance(image, Image.Image)
+
+            # Test not batched input
+            boxes = [[[0, 0, 1, 1], [0.5, 0.5, 0.5, 0.5]]]
+            encoded_images = image_processing(image_inputs[0], boxes=boxes, return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape([image_inputs[0]])
+            self.assertEqual(tuple(encoded_images.shape), (2, *expected_output_image_shape))
+
+            # Test batched
+            boxes = [[[0, 0, 1, 1], [0.5, 0.5, 0.5, 0.5]]] * self.image_processor_tester.batch_size
+            encoded_images = image_processing(image_inputs, boxes=boxes, return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs)
+            self.assertEqual(
+                tuple(encoded_images.shape), (self.image_processor_tester.batch_size * 2, *expected_output_image_shape)
+            )
+
+    def test_call_numpy(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random numpy tensors
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, numpify=True)
+            for image in image_inputs:
+                self.assertIsInstance(image, np.ndarray)
+
+            # Test not batched input
+            boxes = [[[0, 0, 1, 1], [0.5, 0.5, 0.5, 0.5]]]
+            encoded_images = image_processing(image_inputs[0], boxes=boxes, return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape([image_inputs[0]])
+            self.assertEqual(tuple(encoded_images.shape), (2, *expected_output_image_shape))
+
+            # Test batched
+            boxes = [[[0, 0, 1, 1], [0.5, 0.5, 0.5, 0.5]]] * self.image_processor_tester.batch_size
+            encoded_images = image_processing(image_inputs, boxes=boxes, return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs)
+            self.assertEqual(
+                tuple(encoded_images.shape), (self.image_processor_tester.batch_size * 2, *expected_output_image_shape)
+            )
+
+    def test_call_pytorch(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random PyTorch tensors
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
+
+            for image in image_inputs:
+                self.assertIsInstance(image, torch.Tensor)
+
+            # Test not batched input
+            boxes = [[[0, 0, 1, 1], [0.5, 0.5, 0.5, 0.5]]]
+            encoded_images = image_processing(image_inputs[0], boxes=boxes, return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape([image_inputs[0]])
+            self.assertEqual(tuple(encoded_images.shape), (2, *expected_output_image_shape))
+
+            # Test batched
+            boxes = [[[0, 0, 1, 1], [0.5, 0.5, 0.5, 0.5]]] * self.image_processor_tester.batch_size
+            encoded_images = image_processing(image_inputs, boxes=boxes, return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs)
+            self.assertEqual(
+                tuple(encoded_images.shape), (self.image_processor_tester.batch_size * 2, *expected_output_image_shape)
+            )
+
+    def test_call_numpy_4_channels(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processor = image_processing_class(**self.image_processor_dict)
+
+            # create random numpy tensors
+            self.image_processor_tester.num_channels = 4
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, numpify=True)
+            # Test not batched input
+            boxes = [[[0, 0, 1, 1], [0.5, 0.5, 0.5, 0.5]]]
+            encoded_images = image_processor(
+                image_inputs[0],
+                boxes=boxes,
+                return_tensors="pt",
+                input_data_format="channels_last",
+                image_mean=(0.0, 0.0, 0.0, 0.0),
+                image_std=(1.0, 1.0, 1.0, 1.0),
+            ).pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape([image_inputs[0]])
+            self.assertEqual(tuple(encoded_images.shape), (len(boxes[0]), *expected_output_image_shape))
+
+            # Test batched
+            boxes = [[[0, 0, 1, 1], [0.5, 0.5, 0.5, 0.5]]] * self.image_processor_tester.batch_size
+            encoded_images = image_processor(
+                image_inputs,
+                boxes=boxes,
+                return_tensors="pt",
+                input_data_format="channels_last",
+                image_mean=(0.0, 0.0, 0.0, 0.0),
+                image_std=(1.0, 1.0, 1.0, 1.0),
+            ).pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs)
+            self.assertEqual(
+                tuple(encoded_images.shape),
+                (self.image_processor_tester.batch_size * len(boxes[0]), *expected_output_image_shape),
+            )
+            self.image_processor_tester.num_channels = 3
+
+    @require_vision
+    @require_torch
+    def test_backends_equivalence(self):
+        """VitPose requires boxes parameter for preprocessing."""
+        if len(self.image_processing_classes) < 2:
+            self.skipTest(reason="Skipping backends equivalence test as there are less than 2 backends")
+
+        dummy_image = Image.open(
+            io.BytesIO(
+                httpx.get("http://images.cocodataset.org/val2017/000000039769.jpg", follow_redirects=True).content
+            )
+        )
+        boxes = [[[0, 0, 1, 1]]]
+
+        encodings = {}
+        for backend_name, image_processing_class in self.image_processing_classes.items():
+            image_processor = image_processing_class(**self.image_processor_dict)
+            encodings[backend_name] = image_processor(dummy_image, boxes=boxes, return_tensors="pt")
+
+        backend_names = list(encodings.keys())
+        reference_encoding = encodings[backend_names[0]].pixel_values
+        for backend_name in backend_names[1:]:
+            self._assert_tensors_equivalence(reference_encoding, encodings[backend_name].pixel_values)
+
+    @require_vision
+    @require_torch
+    def test_backends_equivalence_batched(self):
+        """VitPose requires boxes parameter for batched preprocessing."""
+        if len(self.image_processing_classes) < 2:
+            self.skipTest(reason="Skipping backends equivalence test as there are less than 2 backends")
+
+        dummy_images = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
+        boxes = [[[0, 0, 1, 1]]] * len(dummy_images)
+
+        encodings = {}
+        for backend_name, image_processing_class in self.image_processing_classes.items():
+            image_processor = image_processing_class(**self.image_processor_dict)
+            encodings[backend_name] = image_processor(dummy_images, boxes=boxes, return_tensors="pt")
+
+        backend_names = list(encodings.keys())
+        reference_encoding = encodings[backend_names[0]].pixel_values
+        for backend_name in backend_names[1:]:
+            self._assert_tensors_equivalence(reference_encoding, encodings[backend_name].pixel_values)
+
+    @require_torch_accelerator
+    @require_vision
+    @pytest.mark.torch_compile_test
+    def test_can_compile_torchvision_backend(self):
+        """VitPose requires boxes parameter for preprocessing."""
+        from transformers.testing_utils import torch_device
+
+        if "torchvision" not in self.image_processing_classes:
+            self.skipTest("Skipping compilation test as torchvision backend is not available")
+
+        torch.compiler.reset()
+        input_image = torch.randint(0, 255, (3, 224, 224), dtype=torch.uint8)
+        image_processor = self.image_processing_classes["torchvision"](**self.image_processor_dict)
+        boxes = [[[0, 0, 1, 1]]]
+        output_eager = image_processor(input_image, boxes=boxes, device=torch_device, return_tensors="pt")
+
+        image_processor = torch.compile(image_processor, mode="reduce-overhead")
+        output_compiled = image_processor(input_image, boxes=boxes, device=torch_device, return_tensors="pt")
+        self._assert_tensors_equivalence(
+            output_eager.pixel_values, output_compiled.pixel_values, atol=1e-4, rtol=1e-4, mean_atol=1e-5
+        )

tests/models/vitpose/test_modeling_vitpose.py

@@ -0,0 +1,329 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Testing suite for the PyTorch VitPose model."""
+
+import inspect
+import unittest
+from functools import cached_property
+
+import requests
+
+from transformers import VitPoseBackboneConfig, VitPoseConfig
+from transformers.testing_utils import require_torch, require_vision, slow, torch_device
+from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import VitPoseForPoseEstimation
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+if is_torchvision_available():
+    from transformers import VitPoseImageProcessor
+
+
+class VitPoseModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        image_size=[16 * 8, 12 * 8],
+        patch_size=[8, 8],
+        num_channels=3,
+        is_training=True,
+        use_labels=True,
+        hidden_size=32,
+        num_hidden_layers=2,
+        num_attention_heads=4,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        type_sequence_label_size=10,
+        initializer_range=0.02,
+        num_labels=2,
+        scale_factor=4,
+        out_indices=[-1],
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.is_training = is_training
+        self.use_labels = use_labels
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.type_sequence_label_size = type_sequence_label_size
+        self.initializer_range = initializer_range
+        self.num_labels = num_labels
+        self.scale_factor = scale_factor
+        self.out_indices = out_indices
+        self.scope = scope
+
+        # in VitPose, the seq length equals the number of patches
+        num_patches = (image_size[0] // patch_size[0]) * (image_size[1] // patch_size[1])
+        self.seq_length = num_patches
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size[0], self.image_size[1]])
+
+        labels = None
+        if self.use_labels:
+            labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+
+        config = self.get_config()
+
+        return config, pixel_values, labels
+
+    def get_config(self):
+        return VitPoseConfig(
+            backbone_config=self.get_backbone_config(),
+        )
+
+    def get_backbone_config(self):
+        return VitPoseBackboneConfig(
+            image_size=self.image_size,
+            patch_size=self.patch_size,
+            num_channels=self.num_channels,
+            num_hidden_layers=self.num_hidden_layers,
+            hidden_size=self.hidden_size,
+            intermediate_size=self.intermediate_size,
+            num_attention_heads=self.num_attention_heads,
+            hidden_act=self.hidden_act,
+            out_indices=self.out_indices,
+        )
+
+    def create_and_check_for_pose_estimation(self, config, pixel_values, labels):
+        model = VitPoseForPoseEstimation(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+
+        expected_height = (self.image_size[0] // self.patch_size[0]) * self.scale_factor
+        expected_width = (self.image_size[1] // self.patch_size[1]) * self.scale_factor
+
+        self.parent.assertEqual(
+            result.heatmaps.shape, (self.batch_size, self.num_labels, expected_height, expected_width)
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            pixel_values,
+            labels,
+        ) = config_and_inputs
+        inputs_dict = {"pixel_values": pixel_values}
+        return config, inputs_dict
+
+
+@require_torch
+class VitPoseModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as VitPose does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (VitPoseForPoseEstimation,) if is_torch_available() else ()
+
+    test_resize_embeddings = False
+
+    def setUp(self):
+        self.model_tester = VitPoseModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=VitPoseConfig, has_text_modality=False, hidden_size=32)
+
+    def test_config(self):
+        self.config_tester.create_and_test_config_to_json_string()
+        self.config_tester.create_and_test_config_to_json_file()
+        self.config_tester.create_and_test_config_from_and_save_pretrained()
+        self.config_tester.create_and_test_config_with_num_labels()
+        self.config_tester.check_config_can_be_init_without_params()
+        self.config_tester.check_config_arguments_init()
+
+    def test_batching_equivalence(self, atol=3e-4, rtol=3e-4):
+        super().test_batching_equivalence(atol=atol, rtol=rtol)
+
+    @unittest.skip(reason="VitPose does not support input and output embeddings")
+    def test_model_common_attributes(self):
+        pass
+
+    @unittest.skip(reason="VitPose does not support input and output embeddings")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="VitPose does not support input and output embeddings")
+    def test_model_get_set_embeddings(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training_gradient_checkpointing_use_reentrant_false(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training_gradient_checkpointing_use_reentrant_true(self):
+        pass
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = ["pixel_values"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    def test_for_pose_estimation(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_pose_estimation(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        model_name = "usyd-community/vitpose-base-simple"
+        model = VitPoseForPoseEstimation.from_pretrained(model_name)
+        self.assertIsNotNone(model)
+
+
+# We will verify our results on an image of people in house
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000000139.jpg"
+    image = Image.open(requests.get(url, stream=True).raw)
+    return image
+
+
+@require_torch
+@require_vision
+class VitPoseModelIntegrationTest(unittest.TestCase):
+    @cached_property
+    def default_image_processor(self):
+        return (
+            VitPoseImageProcessor.from_pretrained("usyd-community/vitpose-base-simple")
+            if is_vision_available()
+            else None
+        )
+
+    @slow
+    def test_inference_pose_estimation(self):
+        image_processor = self.default_image_processor
+        model = VitPoseForPoseEstimation.from_pretrained("usyd-community/vitpose-base-simple", device_map=torch_device)
+
+        image = prepare_img()
+        boxes = [[[412.8, 157.61, 53.05, 138.01], [384.43, 172.21, 15.12, 35.74]]]
+
+        inputs = image_processor(images=image, boxes=boxes, return_tensors="pt").to(torch_device)
+
+        with torch.no_grad():
+            outputs = model(**inputs)
+        heatmaps = outputs.heatmaps
+
+        assert heatmaps.shape == (2, 17, 64, 48)
+
+        expected_slice = torch.tensor(
+            [
+                [9.9330e-06, 9.9330e-06, 9.9330e-06],
+                [9.9330e-06, 9.9330e-06, 9.9330e-06],
+                [9.9330e-06, 9.9330e-06, 9.9330e-06],
+            ]
+        ).to(torch_device)
+
+        assert torch.allclose(heatmaps[0, 0, :3, :3], expected_slice, atol=1e-4)
+
+        pose_results = image_processor.post_process_pose_estimation(outputs, boxes=boxes)[0]
+
+        expected_bbox = torch.tensor([391.9900, 190.0800, 391.1575, 189.3034])
+        expected_keypoints = torch.tensor(
+            [
+                [3.9813e02, 1.8184e02],
+                [3.9828e02, 1.7981e02],
+                [3.9596e02, 1.7948e02],
+            ]
+        )
+        expected_scores = torch.tensor([8.7529e-01, 8.4315e-01, 9.2678e-01])
+
+        self.assertEqual(len(pose_results), 2)
+        torch.testing.assert_close(pose_results[1]["bbox"].cpu(), expected_bbox, rtol=1e-4, atol=1e-4)
+        torch.testing.assert_close(pose_results[1]["keypoints"][:3].cpu(), expected_keypoints, rtol=1e-2, atol=1e-2)
+        torch.testing.assert_close(pose_results[1]["scores"][:3].cpu(), expected_scores, rtol=1e-4, atol=1e-4)
+
+    @slow
+    def test_batched_inference(self):
+        image_processor = self.default_image_processor
+        model = VitPoseForPoseEstimation.from_pretrained("usyd-community/vitpose-base-simple", device_map=torch_device)
+
+        image = prepare_img()
+        boxes = [
+            [[412.8, 157.61, 53.05, 138.01], [384.43, 172.21, 15.12, 35.74]],
+            [[412.8, 157.61, 53.05, 138.01], [384.43, 172.21, 15.12, 35.74]],
+        ]
+
+        inputs = image_processor(images=[image, image], boxes=boxes, return_tensors="pt").to(torch_device)
+
+        with torch.no_grad():
+            outputs = model(**inputs)
+        heatmaps = outputs.heatmaps
+
+        assert heatmaps.shape == (4, 17, 64, 48)
+
+        expected_slice = torch.tensor(
+            [
+                [9.9330e-06, 9.9330e-06, 9.9330e-06],
+                [9.9330e-06, 9.9330e-06, 9.9330e-06],
+                [9.9330e-06, 9.9330e-06, 9.9330e-06],
+            ]
+        ).to(torch_device)
+
+        assert torch.allclose(heatmaps[0, 0, :3, :3], expected_slice, atol=1e-4)
+
+        pose_results = image_processor.post_process_pose_estimation(outputs, boxes=boxes)
+
+        expected_bbox = torch.tensor([391.9900, 190.0800, 391.1575, 189.3034])
+        expected_keypoints = torch.tensor(
+            [
+                [3.9813e02, 1.8184e02],
+                [3.9828e02, 1.7981e02],
+                [3.9596e02, 1.7948e02],
+            ]
+        )
+        expected_scores = torch.tensor([8.7529e-01, 8.4315e-01, 9.2678e-01])
+
+        self.assertEqual(len(pose_results), 2)
+        self.assertEqual(len(pose_results[0]), 2)
+        torch.testing.assert_close(pose_results[0][1]["bbox"].cpu(), expected_bbox, rtol=1e-4, atol=1e-4)
+        torch.testing.assert_close(pose_results[0][1]["keypoints"][:3].cpu(), expected_keypoints, rtol=1e-2, atol=1e-2)
+        torch.testing.assert_close(pose_results[0][1]["scores"][:3].cpu(), expected_scores, rtol=1e-4, atol=1e-4)