#!/usr/bin/env python3 """Visualize augmentation ranges that simulate viewpoint changes.""" import cv2 import numpy as np from pathlib import Path # Load a sample frame frame_path = "/data/libero/ood_objpos_v3/libero_spatial/task_0/demo_0/frames/000005.png" img = cv2.imread(frame_path) H, W = img.shape[:2] # 448x448 n_cols = 8 # number of samples per augmentation # ============================================================ # Define augmentation dimensions with min/max ranges # ============================================================ def apply_crop(img, dx, dy, crop_frac=0.85): """Simulate lateral camera shift via off-center crop + resize. dx, dy in pixels: how far the crop center shifts from image center. crop_frac: fraction of image to keep (smaller = more zoom). """ H, W = img.shape[:2] crop_h, crop_w = int(H * crop_frac), int(W * crop_frac) cx, cy = W // 2 + int(dx), H // 2 + int(dy) x1 = max(0, min(cx - crop_w // 2, W - crop_w)) y1 = max(0, min(cy - crop_h // 2, H - crop_h)) cropped = img[y1:y1+crop_h, x1:x1+crop_w] return cv2.resize(cropped, (W, H)) def apply_rotation(img, angle_deg): """In-plane rotation around image center.""" H, W = img.shape[:2] M = cv2.getRotationMatrix2D((W/2, H/2), angle_deg, 1.0) return cv2.warpAffine(img, M, (W, H), borderMode=cv2.BORDER_REFLECT_101) def apply_hshear(img, shear_x): """Horizontal shear — simulates horizontal oblique viewing.""" H, W = img.shape[:2] M = np.float32([[1, shear_x, -shear_x * H/2], [0, 1, 0]]) return cv2.warpAffine(img, M, (W, H), borderMode=cv2.BORDER_REFLECT_101) def apply_vshear(img, shear_y): """Vertical shear — simulates vertical oblique viewing (elevation change).""" H, W = img.shape[:2] M = np.float32([[1, 0, 0], [shear_y, 1, -shear_y * W/2]]) return cv2.warpAffine(img, M, (W, H), borderMode=cv2.BORDER_REFLECT_101) def apply_perspective(img, strength, direction='horizontal'): """Perspective warp — more realistic viewpoint simulation than shear. strength: how strong the perspective effect is (0 = none, 0.15 = strong). """ H, W = img.shape[:2] src = np.float32([[0, 0], [W, 0], [W, H], [0, H]]) s = strength if direction == 'horizontal': # Simulate looking from the left (positive) or right (negative) dst = np.float32([[0 + s*W, 0 + s*H/3], [W - s*W, 0 - s*H/3], [W + s*W, H + s*H/3], [0 - s*W, H - s*H/3]]) else: # vertical # Simulate looking from above (positive) or below (negative) dst = np.float32([[0 + s*W/3, 0 + s*H], [W - s*W/3, 0 + s*H], [W + s*W/3, H - s*H], [0 - s*W/3, H - s*H]]) M = cv2.getPerspectiveTransform(src, dst) return cv2.warpPerspective(img, M, (W, H), borderMode=cv2.BORDER_REFLECT_101) # ============================================================ # Augmentation specs: (name, function, param_values, param_label) # ============================================================ augmentations = [ ("Horizontal Crop\n(lateral shift)", lambda img, v: apply_crop(img, dx=v, dy=0, crop_frac=0.82), np.linspace(-50, 50, n_cols), "dx (px)"), ("Vertical Crop\n(elevation shift)", lambda img, v: apply_crop(img, dx=0, dy=v, crop_frac=0.82), np.linspace(-50, 50, n_cols), "dy (px)"), ("Rotation\n(camera roll)", lambda img, v: apply_rotation(img, v), np.linspace(-15, 15, n_cols), "angle (deg)"), ("Horizontal Shear\n(azimuth sim)", lambda img, v: apply_hshear(img, v), np.linspace(-0.20, 0.20, n_cols), "shear_x"), ("Vertical Shear\n(elevation sim)", lambda img, v: apply_vshear(img, v), np.linspace(-0.20, 0.20, n_cols), "shear_y"), ("Horizontal Perspective\n(azimuth viewpoint)", lambda img, v: apply_perspective(img, v, 'horizontal'), np.linspace(-0.15, 0.15, n_cols), "strength"), ("Vertical Perspective\n(elevation viewpoint)", lambda img, v: apply_perspective(img, v, 'vertical'), np.linspace(-0.15, 0.15, n_cols), "strength"), ] # ============================================================ # Build the grid image # ============================================================ thumb = 160 row_label_w = 180 col_label_h = 30 pad = 3 grid_w = row_label_w + n_cols * (thumb + pad) grid_h = col_label_h + len(augmentations) * (thumb + pad + 18) # 18 for param label canvas = np.zeros((grid_h, grid_w, 3), dtype=np.uint8) canvas[:] = (30, 30, 30) y = col_label_h for aug_name, aug_fn, params, param_label in augmentations: # Row label lines = aug_name.split('\n') for li, line in enumerate(lines): cv2.putText(canvas, line, (8, y + thumb//2 - 10 + li*18), cv2.FONT_HERSHEY_SIMPLEX, 0.42, (200, 200, 200), 1) # Generate augmented images x = row_label_w for i, val in enumerate(params): aug_img = aug_fn(img.copy(), val) aug_thumb = cv2.resize(aug_img, (thumb, thumb)) # Highlight center (zero/neutral) column is_center = (i == n_cols // 2 - 1) or (i == n_cols // 2) if abs(val) < 1e-6 or (abs(val) == min(abs(params))): # Closest to zero pass # Border color: blue for min, green for center, red for max t = i / (n_cols - 1) # 0 to 1 if t < 0.5: # Blue to white color = (255, int(200*t*2), int(100*t*2)) else: # White to red color = (int(255*(1-(t-0.5)*2)), int(200*(1-(t-0.5)*2)), 255) aug_thumb[:2, :] = color; aug_thumb[-2:, :] = color aug_thumb[:, :2] = color; aug_thumb[:, -2:] = color canvas[y:y+thumb, x:x+thumb] = aug_thumb # Parameter value label below lbl = f"{val:.1f}" if abs(val) >= 1 else f"{val:.3f}" cv2.putText(canvas, lbl, (x + 2, y + thumb + 13), cv2.FONT_HERSHEY_SIMPLEX, 0.33, (150, 150, 150), 1) x += thumb + pad y += thumb + pad + 18 # Title cv2.putText(canvas, "Augmentation Range Exploration (viewpoint simulation)", (row_label_w, 20), cv2.FONT_HERSHEY_SIMPLEX, 0.55, (100, 200, 255), 1) out_path = "results/augmentation_grid.png" cv2.imwrite(out_path, canvas) print(f"Saved: {out_path} ({canvas.shape[1]}x{canvas.shape[0]})")