180 lines
6.3 KiB
Python
180 lines
6.3 KiB
Python
from pathlib import Path
|
|
import numpy as np
|
|
from PIL import Image, ImageDraw
|
|
import matplotlib.pyplot as plt
|
|
from matplotlib.tri import Triangulation
|
|
|
|
from image_preparing import fix_image_resolution
|
|
|
|
import time
|
|
|
|
|
|
transformation_resolution = (1600, 1200)
|
|
|
|
def auto_correspondences_grid(pil, nx=8, ny=6):
|
|
w, h = pil.size
|
|
xs = np.linspace(0, w - 1, nx)
|
|
ys = np.linspace(0, h - 1, ny)
|
|
pts = []
|
|
for y in ys:
|
|
for x in xs:
|
|
pts.append([float(x), float(y)])
|
|
corners = [[0.0, 0.0], [w - 1.0, 0.0], [w - 1.0, h - 1.0], [0.0, h - 1.0]]
|
|
for c in corners:
|
|
if c not in pts:
|
|
pts.insert(0, c)
|
|
return np.array(pts, dtype=float), np.array(pts, dtype=float).copy()
|
|
def compute_delaunay_on_points(pts):
|
|
pts = np.asarray(pts)
|
|
tri = Triangulation(pts[:, 0], pts[:, 1])
|
|
return tri.triangles
|
|
def _affine_coeffs(from_tri, to_tri):
|
|
A = []
|
|
B = []
|
|
for (x_dst, y_dst), (x_src, y_src) in zip(from_tri, to_tri):
|
|
A.append([x_dst, y_dst, 1, 0, 0, 0])
|
|
A.append([0, 0, 0, x_dst, y_dst, 1])
|
|
B.append(x_src)
|
|
B.append(y_src)
|
|
sol, *_ = np.linalg.lstsq(np.array(A), np.array(B), rcond=None)
|
|
return tuple(sol.tolist())
|
|
def _warp_triangle(src_pil, src_tri, dst_tri, out_size, offset):
|
|
dx, dy = offset
|
|
dst_local = np.array(dst_tri) - np.array([dx, dy])
|
|
coeffs = _affine_coeffs(dst_local, src_tri)
|
|
return src_pil.transform(out_size, Image.AFFINE, coeffs, resample=Image.BILINEAR)
|
|
def reconstruct_goal_from_source(src_pil, pts1, pts2):
|
|
w, h = src_pil.size
|
|
pts1 = np.asarray(pts1)
|
|
pts2 = np.asarray(pts2)
|
|
triangles = compute_delaunay_on_points(pts2)
|
|
out = Image.new("RGB", (w, h))
|
|
for tri in triangles:
|
|
tri = np.asarray(tri, dtype=int)
|
|
src_tri = pts1[tri]
|
|
dst_tri = pts2[tri]
|
|
min_x = max(int(np.floor(dst_tri[:, 0].min())), 0)
|
|
max_x = min(int(np.ceil(dst_tri[:, 0].max())), w)
|
|
min_y = max(int(np.floor(dst_tri[:, 1].min())), 0)
|
|
max_y = min(int(np.ceil(dst_tri[:, 1].max())), h)
|
|
if max_x <= min_x or max_y <= min_y:
|
|
continue
|
|
out_w = max_x - min_x
|
|
out_h = max_y - min_y
|
|
warped = _warp_triangle(src_pil, src_tri, dst_tri, (out_w, out_h), (min_x, min_y))
|
|
mask = Image.new("L", (out_w, out_h), 0)
|
|
draw = ImageDraw.Draw(mask)
|
|
tri_local = [(x - min_x, y - min_y) for x, y in dst_tri.tolist()]
|
|
draw.polygon(tri_local, fill=255)
|
|
out.paste(warped, (min_x, min_y), mask)
|
|
return out
|
|
def transform_images(src_path, goal_path):
|
|
src = fix_image_resolution(src_path, size=transformation_resolution)
|
|
goal = fix_image_resolution(goal_path, size=transformation_resolution)
|
|
start = time.time()
|
|
image = tile_reposition_transition(src, goal, grid=transformation_resolution)
|
|
end = time.time()
|
|
print(f"Transition took {end - start:.2f} seconds")
|
|
fig, ax = plt.subplots(figsize=(8, 6), facecolor="black")
|
|
fig.patch.set_facecolor("black")
|
|
ax.set_facecolor("black")
|
|
ax.imshow(image, cmap=None)
|
|
ax.axis("off")
|
|
|
|
def _close_on_key(event):
|
|
if event.key in ("q", "escape"):
|
|
plt.close(fig)
|
|
|
|
fig.canvas.mpl_connect("key_press_event", _close_on_key)
|
|
|
|
plt.show()
|
|
def _grid_edges(size, divisions):
|
|
return np.rint(np.linspace(0, size, divisions + 1)).astype(np.int32)
|
|
def _build_tile_boxes_and_centers(w, h, nx, ny):
|
|
x_edges = _grid_edges(w, nx)
|
|
y_edges = _grid_edges(h, ny)
|
|
|
|
x0 = np.repeat(x_edges[:-1], ny)
|
|
x1 = np.repeat(x_edges[1:], ny)
|
|
y0 = np.tile(y_edges[:-1], nx)
|
|
y1 = np.tile(y_edges[1:], nx)
|
|
|
|
boxes = np.stack([x0, y0, x1, y1], axis=1)
|
|
centers = np.stack([(x0 + x1) * 0.5, (y0 + y1) * 0.5], axis=1).astype(np.float32)
|
|
areas = np.maximum((x1 - x0) * (y1 - y0), 1).astype(np.float32)
|
|
return boxes, centers, areas
|
|
def _integral_image_rgb(arr):
|
|
integral = arr.cumsum(axis=0).cumsum(axis=1)
|
|
return np.pad(integral, ((1, 0), (1, 0), (0, 0)), mode="constant", constant_values=0)
|
|
def _tile_means_from_integral(integral, boxes, areas):
|
|
x0 = boxes[:, 0]
|
|
y0 = boxes[:, 1]
|
|
x1 = boxes[:, 2]
|
|
y1 = boxes[:, 3]
|
|
|
|
sums = (
|
|
integral[y1, x1]
|
|
- integral[y0, x1]
|
|
- integral[y1, x0]
|
|
+ integral[y0, x0]
|
|
)
|
|
return sums / areas[:, None]
|
|
def _match_tiles_by_feature_rank(src_avgs, src_centers, goal_avgs, goal_centers, w, h):
|
|
color_scale = 255.0 * np.sqrt(3.0)
|
|
spatial_scale = max(np.hypot(w, h), 1e-6)
|
|
|
|
src_feat = np.hstack([
|
|
src_avgs / color_scale,
|
|
(src_centers / spatial_scale) * 0.35,
|
|
])
|
|
goal_feat = np.hstack([
|
|
goal_avgs / color_scale,
|
|
(goal_centers / spatial_scale) * 0.35,
|
|
])
|
|
|
|
# Fixed projection keeps pairing deterministic and avoids O(n^2) matching.
|
|
projection = np.array([0.50, 0.30, 0.20, 0.60, 0.40], dtype=np.float32)
|
|
src_order = np.argsort(src_feat @ projection)
|
|
goal_order = np.argsort(goal_feat @ projection)
|
|
return src_order, goal_order
|
|
def tile_reposition_transition(src_pil, goal_pil, grid=transformation_resolution):
|
|
src = src_pil.convert("RGB")
|
|
goal = goal_pil.convert("RGB")
|
|
w, h = src.size
|
|
nx, ny = grid
|
|
|
|
src_boxes, src_centers, src_areas = _build_tile_boxes_and_centers(w, h, nx, ny)
|
|
goal_boxes, goal_centers, goal_areas = _build_tile_boxes_and_centers(w, h, nx, ny)
|
|
|
|
src_arr = np.asarray(src, dtype=np.float32)
|
|
goal_arr = np.asarray(goal, dtype=np.float32)
|
|
src_integral = _integral_image_rgb(src_arr)
|
|
goal_integral = _integral_image_rgb(goal_arr)
|
|
|
|
src_avgs = _tile_means_from_integral(src_integral, src_boxes, src_areas)
|
|
goal_avgs = _tile_means_from_integral(goal_integral, goal_boxes, goal_areas)
|
|
|
|
src_order, goal_order = _match_tiles_by_feature_rank(
|
|
src_avgs,
|
|
src_centers,
|
|
goal_avgs,
|
|
goal_centers,
|
|
w,
|
|
h,
|
|
)
|
|
|
|
final = Image.new("RGB", (w, h))
|
|
for s_idx, g_idx in zip(src_order.tolist(), goal_order.tolist()):
|
|
sx0, sy0, sx1, sy1 = src_boxes[s_idx].tolist()
|
|
gx0, gy0, gx1, gy1 = goal_boxes[g_idx].tolist()
|
|
|
|
tile = src.crop((sx0, sy0, sx1, sy1))
|
|
goal_w = max(gx1 - gx0, 1)
|
|
goal_h = max(gy1 - gy0, 1)
|
|
if tile.size != (goal_w, goal_h):
|
|
tile = tile.resize((goal_w, goal_h), Image.Resampling.BILINEAR)
|
|
final.paste(tile, (gx0, gy0))
|
|
return final
|
|
|
|
|