import numpy as np import pytest from vla_foundry.data.preprocessing.image_utils import ImageResizingMethod from vla_foundry.data.robotics.cv_utils import scale_intrinsics_4_for_resize_and_crop def assert_allclose(actual, expected, rtol=1e-6, atol=1e-6): np.testing.assert_allclose(np.asarray(actual), np.asarray(expected), rtol=rtol, atol=atol) class TestScaleIntrinsics4ForResizeAndCrop: """Tests for scale_intrinsics_4_for_resize_and_crop (fx, fy, cx, cy).""" @pytest.fixture def sample_intrinsics_4(self): # fx, fy, cx, cy return np.array([800.0, 800.0, 320.0, 240.0], dtype=np.float64) @pytest.fixture def batch_intrinsics_4(self): return np.array( [ [800.0, 800.0, 320.0, 240.0], [1000.0, 1000.0, 512.0, 384.0], ], dtype=np.float64, ) # ------------------------- # RESIZE_NO_CROP (distort) # ------------------------- def test_resize_no_crop_single(self, sample_intrinsics_4): original_size = (640, 480) processed_size = (320, 240) # exact 0.5 in both result = scale_intrinsics_4_for_resize_and_crop( sample_intrinsics_4, original_size, processed_size, resize_method=ImageResizingMethod.RESIZE_NO_CROP, ) expected = np.array([400.0, 400.0, 160.0, 120.0], dtype=np.float64) assert_allclose(result, expected) def test_resize_no_crop_batch(self, batch_intrinsics_4): original_size = (1024, 768) processed_size = (512, 384) # exact 0.5 in both result = scale_intrinsics_4_for_resize_and_crop( batch_intrinsics_4, original_size, processed_size, resize_method=ImageResizingMethod.RESIZE_NO_CROP, ) expected = np.array( [ [400.0, 400.0, 160.0, 120.0], [500.0, 500.0, 256.0, 192.0], ], dtype=np.float64, ) assert_allclose(result, expected) def test_resize_no_crop_anisotropic(self, sample_intrinsics_4): # Different scale factors in x and y original_size = (640, 480) processed_size = (320, 480) # sx=0.5, sy=1.0 result = scale_intrinsics_4_for_resize_and_crop( sample_intrinsics_4, original_size, processed_size, resize_method=ImageResizingMethod.RESIZE_NO_CROP, ) expected = np.array( [ 800.0 * 0.5, # fx 800.0 * 1.0, # fy 320.0 * 0.5, # cx 240.0 * 1.0, # cy ], dtype=np.float64, ) assert_allclose(result, expected) # ------------------------- # CENTER_CROP (cover + crop) # ------------------------- def test_center_crop_square_from_landscape(self, sample_intrinsics_4): # original 640x480 -> target 480x480 # scale = max(480/640=0.75, 480/480=1.0) = 1.0 # new_W=640,new_H=480; cx_offset=(640-480)/2=80; cy_offset=(480-480)/2=0 original_size = (640, 480) processed_size = (480, 480) result = scale_intrinsics_4_for_resize_and_crop( sample_intrinsics_4, original_size, processed_size, resize_method=ImageResizingMethod.CENTER_CROP, ) expected = np.array( [ 800.0 * 1.0, # fx 800.0 * 1.0, # fy 320.0 * 1.0 - 80.0, # cx 240.0 * 1.0 - 0.0, # cy ], dtype=np.float64, ) assert_allclose(result, expected) def test_center_crop_square_smaller(self, sample_intrinsics_4): # original 640x480 -> target 320x320 # scale = max(320/640=0.5, 320/480=0.666666...) = 2/3 # new_W=640*(2/3)=426.666..., new_H=480*(2/3)=320 # cx_offset=(426.666...-320)/2=53.333..., cy_offset=(320-320)/2=0 original_size = (640, 480) processed_size = (320, 320) scale = 320.0 / 480.0 # 2/3 new_W = 640.0 * scale new_H = 480.0 * scale cx_offset = (new_W - 320.0) / 2.0 cy_offset = (new_H - 320.0) / 2.0 result = scale_intrinsics_4_for_resize_and_crop( sample_intrinsics_4, original_size, processed_size, resize_method=ImageResizingMethod.CENTER_CROP, ) expected = np.array( [ 800.0 * scale, 800.0 * scale, 320.0 * scale - cx_offset, 240.0 * scale - cy_offset, ], dtype=np.float64, ) assert_allclose(result, expected, rtol=1e-6, atol=1e-6) def test_center_crop_batch(self, batch_intrinsics_4): original_size = (640, 480) processed_size = (480, 480) # scale = 1.0; cx_offset=80; cy_offset=0 result = scale_intrinsics_4_for_resize_and_crop( batch_intrinsics_4, original_size, processed_size, resize_method=ImageResizingMethod.CENTER_CROP, ) expected = np.array( [ [800.0, 800.0, 320.0 - 80.0, 240.0], [1000.0, 1000.0, 512.0 - 80.0, 384.0], ], dtype=np.float64, ) assert_allclose(result, expected) # ------------------------- # RESIZE_FIT (fit + pad) # ------------------------- def test_resize_fit_letterbox_padding(self, sample_intrinsics_4): # original 640x480 -> target 640x640 (pad top/bottom) # scale = min(640/640=1, 640/480=1.333...) = 1 # new_W=640, new_H=480 # pad_x=(640-640)/2=0, pad_y=(640-480)/2=80 original_size = (640, 480) processed_size = (640, 640) result = scale_intrinsics_4_for_resize_and_crop( sample_intrinsics_4, original_size, processed_size, resize_method=ImageResizingMethod.RESIZE_FIT, ) expected = np.array( [ 800.0 * 1.0, 800.0 * 1.0, 320.0 * 1.0 + 0.0, 240.0 * 1.0 + 80.0, ], dtype=np.float64, ) assert_allclose(result, expected) def test_resize_fit_pillarbox_padding(self, sample_intrinsics_4): # original 640x480 -> target 800x480 (pad left/right) # scale=min(800/640=1.25, 480/480=1) = 1 # new_W=640, new_H=480 # pad_x=(800-640)/2=80, pad_y=0 original_size = (640, 480) processed_size = (800, 480) result = scale_intrinsics_4_for_resize_and_crop( sample_intrinsics_4, original_size, processed_size, resize_method=ImageResizingMethod.RESIZE_FIT, ) expected = np.array( [ 800.0, 800.0, 320.0 + 80.0, 240.0, ], dtype=np.float64, ) assert_allclose(result, expected) def test_resize_fit_downscale_with_padding(self, sample_intrinsics_4): # original 640x480 -> target 320x320 (fit inside -> pad y) # scale = min(320/640=0.5, 320/480=0.666...) = 0.5 # new_W=320, new_H=240 # pad_x=0, pad_y=(320-240)/2=40 original_size = (640, 480) processed_size = (320, 320) scale = 0.5 pad_x = 0.0 pad_y = (320.0 - 240.0) / 2.0 # 40 result = scale_intrinsics_4_for_resize_and_crop( sample_intrinsics_4, original_size, processed_size, resize_method=ImageResizingMethod.RESIZE_FIT, ) expected = np.array( [ 800.0 * scale, 800.0 * scale, 320.0 * scale + pad_x, 240.0 * scale + pad_y, ], dtype=np.float64, ) assert_allclose(result, expected) # ------------------------- # Basic properties / edge cases # ------------------------- def test_same_size_identity_for_all_methods(self, sample_intrinsics_4): original_size = (640, 480) processed_size = (640, 480) for method in [ ImageResizingMethod.CENTER_CROP, ImageResizingMethod.RESIZE_NO_CROP, ImageResizingMethod.RESIZE_FIT, ]: result = scale_intrinsics_4_for_resize_and_crop( sample_intrinsics_4, original_size, processed_size, resize_method=method, ) assert_allclose(result, sample_intrinsics_4) def test_output_shape_single(self, sample_intrinsics_4): result = scale_intrinsics_4_for_resize_and_crop( sample_intrinsics_4, (640, 480), (320, 240), resize_method=ImageResizingMethod.RESIZE_NO_CROP ) assert np.asarray(result).shape == (4,) def test_output_shape_batch(self, batch_intrinsics_4): result = scale_intrinsics_4_for_resize_and_crop( batch_intrinsics_4, (640, 480), (320, 240), resize_method=ImageResizingMethod.RESIZE_NO_CROP ) assert np.asarray(result).shape == (2, 4) def test_invalid_resize_method_raises(self, sample_intrinsics_4): with pytest.raises(ValueError): scale_intrinsics_4_for_resize_and_crop( sample_intrinsics_4, (640, 480), (320, 240), resize_method="invalid_method", ) def test_wrong_intrinsics_ndim_raises(self): intrinsics = np.zeros((2, 2, 4), dtype=np.float64) # ndim=3 with pytest.raises(ValueError): scale_intrinsics_4_for_resize_and_crop( intrinsics, (640, 480), (320, 240), resize_method=ImageResizingMethod.RESIZE_NO_CROP, ) def test_wrong_intrinsics_last_dim_raises(self): intrinsics = np.zeros((3,), dtype=np.float64) # last dim != 4 with pytest.raises(ValueError): scale_intrinsics_4_for_resize_and_crop( intrinsics, (640, 480), (320, 240), resize_method=ImageResizingMethod.RESIZE_NO_CROP, ) intrinsics2 = np.zeros((2, 5), dtype=np.float64) # last dim != 4 with pytest.raises(ValueError): scale_intrinsics_4_for_resize_and_crop( intrinsics2, (640, 480), (320, 240), resize_method=ImageResizingMethod.RESIZE_NO_CROP, ) def test_zero_or_negative_sizes_do_not_nan(self, sample_intrinsics_4): # Your function doesn't explicitly validate sizes; at minimum ensure it doesn't produce NaNs/Infs # (If you later add validation, change this test to expect ValueError.) for processed_size in [(0, 240), (320, 0), (-320, 240), (320, -240)]: result = scale_intrinsics_4_for_resize_and_crop( sample_intrinsics_4, (640, 480), processed_size, resize_method=ImageResizingMethod.RESIZE_NO_CROP, ) assert np.isfinite(np.asarray(result)).all() def test_batch_consistency(self, batch_intrinsics_4): original_size = (640, 480) processed_size = (320, 240) batch_result = scale_intrinsics_4_for_resize_and_crop( batch_intrinsics_4, original_size, processed_size, resize_method=ImageResizingMethod.RESIZE_NO_CROP, ) indiv = [] for i in range(batch_intrinsics_4.shape[0]): indiv.append( scale_intrinsics_4_for_resize_and_crop( batch_intrinsics_4[i], original_size, processed_size, resize_method=ImageResizingMethod.RESIZE_NO_CROP, ) ) indiv = np.stack(indiv, axis=0) assert_allclose(batch_result, indiv) def test_dtype_is_float(self, sample_intrinsics_4): for dtype in [np.float32, np.float64]: intr = sample_intrinsics_4.astype(dtype) result = scale_intrinsics_4_for_resize_and_crop( intr, (640, 480), (320, 240), resize_method=ImageResizingMethod.RESIZE_FIT ) assert np.issubdtype(np.asarray(result).dtype, np.floating) def test_numerical_precision_exact_half_scale(self): intrinsics = np.array([1000.0, 1000.0, 500.0, 400.0], dtype=np.float64) original_size = (1000, 800) processed_size = (500, 400) # exact 0.5 for both axes result = scale_intrinsics_4_for_resize_and_crop( intrinsics, original_size, processed_size, resize_method=ImageResizingMethod.RESIZE_NO_CROP, ) expected = np.array([500.0, 500.0, 250.0, 200.0], dtype=np.float64) assert_allclose(result, expected, rtol=1e-12, atol=1e-12)