def compute_distance(df, image, th=0.92, label="Parking Slots", plot=False):
df.reset_index(drop=True, inplace=True)
n = len(df)
base_col = ['x1', 'y1', 'x2', 'y2', 'xc', 'yc', 'w', 'b']
df.reset_index(drop=True, inplace=True)
mat, _, _, _ = assign_next_frame(df, df, th=0.6)
np.fill_diagonal(mat, -9)
mat = np.tril(mat)
count = n
to_merge = []
while count > 0:
r, k = np.unravel_index(np.argmax(mat, axis=None), mat.shape)
if mat[r, k] > th:
to_merge.append([r, k])
mat[r, :] = -9
mat[:, k] = -9
mat[k, :] = -9
mat[:, r] = -9
count = count - 1
# print(to_merge)
for i in range(len(to_merge)):
r = to_merge[i][0]
k = to_merge[i][1]
df.loc[r, base_col] = (df.loc[r, base_col] * df.loc[r, "found"] +
df.loc[r, base_col] * df.loc[k, "found"]) / (
df.loc[r, "found"] + df.loc[k, "found"])
df.at[r, "found"] = df.at[r, "found"] + df.at[k, "found"]
df.drop(k, axis=0, inplace=True)
if plot:
visualize.display_instances(image,
df[["y1", "x1", "y2", "x2"]].values,
None,
df["class"].values,
class_names,
scores=None,
title=label,
figsize=(16, 16),
ax=None,
show_mask=False,
show_bbox=True,
colors=None,
captions=df["labels"].astype(str).values)
return df
def compute_distance(df, image, th = 0.92, label = "Parking Slots", plot = False):
df.reset_index(drop=True, inplace=True)
n = len(df)
base_col = ['x1', 'y1', 'x2', 'y2', 'xc', 'yc', 'w' , 'b']
df.reset_index(drop=True, inplace=True)
mat, _, _, _ = assign_next_frame(df, df, th = 0.6)
np.fill_diagonal(mat, -9)
mat = np.tril(mat)
count = n
to_merge = []
while count > 0:
r,k = np.unravel_index(np.argmax(mat, axis=None), mat.shape)
if mat[r,k] > th :
to_merge.append([r,k])
mat[r,:] = -9
mat[:,k] = -9
mat[k,:] = -9
mat[:,r] = -9
count = count -1
# print(to_merge)
for i in range(len(to_merge)):
r = to_merge[i][0]
k = to_merge[i][1]
df.loc[r,base_col] =(df.loc[r,base_col] * df.loc[r,"found"] + df.loc[r,base_col]* df.loc[k,"found"])/(df.loc[r,"found"]+df.loc[k,"found"])
df.at[r,"found"] = df.at[r,"found"]+ df.at[k,"found"]
df.drop(k, axis=0, inplace = True)
if plot :
plt.figure()
box_img = yolo.draw_rect(PILImage.fromarray(cv2.imread(image)),\
df[["y1","x2","y2","x1"]].values, df.index.values, df["class"], df["labels"].values)
f = plt.figure(figsize=(20,40))
plt.imshow(box_img)
plt.title(label)
plt.show()
plt.close()
return df
def look_for_slots(data,
img=[],
PRUNE_TH=3,
plot=True,
PRUNE_STEP=10,
MERGE_STEP=20,
MERGE_TH=0.65):
n_fr = data["frame"].nunique()
cols = [
"labels", 'x1', 'y1', 'x2', 'y2', 'xc', 'yc', 'w', 'b', "class", 'a'
]
base_col = ['x1', 'y1', 'x2', 'y2', 'xc', 'yc', 'w', 'b']
slots = data[data["frame"] == 0][cols]
slots["found"] = 1
def plot_images():
df = slots[['x1', 'y1', 'x2', 'y2', "found", "labels"]]
df["class"] = "empty"
msk = df["labels"].isin(id_map.values())
df.loc[msk, "class"] = "occupy"
nw_df = new[['x1', 'y1', 'x2', 'y2', "labels"]]
nw_df["class"] = "new"
nw_df["found"] = 1
df = df.append(nw_df, sort=True).reset_index(drop=True)
df["found"] = df["found"].astype(int)
plot_frame(img[i], df[["y1", "x2", "y2", "x1"]].values,
df["class"].values, df["found"], df["labels"])
# out_boxes, out_classes, found, labels
# "empty":"#4a148c","occupy":"#f44336", "new":"#7cb342","del":"#80deea"
print("LOOKING FOR PARKING SLOTS INSIDE IMAGE FRAMES")
for i in tqdm(range(1, n_fr)):
post = data[data["frame"] == i].reset_index(drop=True)
_, iou, id_map, status = assign_next_frame(slots, post, th=MERGE_TH)
#print(id_map.keys(), status.sum())
## found again
mask = post["labels"].isin(id_map.keys())
slots.loc[status, "found"] = slots.loc[status, "found"] + 1
occupy = post[mask]
occupy["labels"] = occupy["labels"].map(id_map)
slots.sort_values(by=["labels"], inplace=True)
slots.reset_index(drop=True, inplace=True)
occupy.sort_values(by=["labels"], inplace=True)
occupy.reset_index(drop=True, inplace=True)
slots.loc[status, base_col] = slots.loc[status, base_col].values * (
1 - 1 / (i + 1)) + occupy[base_col].values / (i + 1)
# clean up
if i % PRUNE_STEP == 0:
slots.drop(slots[slots["found"] < PRUNE_TH + 1].index,
inplace=True)
#print(slots)
# merge
if i % MERGE_STEP == 0:
slots = compute_distance(slots,
img[i - 1],
th=MERGE_TH,
label="Parking Slots " + str(i))
# new
idx = np.logical_not(post["labels"].isin(id_map.keys()))
new = post[idx]
new["labels"] = new["labels"] + slots["labels"].max() + 1
new = new[cols]
if len(new) > 0:
new["found"] = 1
slots = slots.append(new, sort=True).reset_index(drop=True)
if plot and (i % MERGE_STEP * 5 == 0):
plot_images()
slots = compute_distance(slots,
img[0],
th=MERGE_TH * 0.9,
label="Parking Slots " + str(MERGE_STEP))
slots.drop(slots[slots["found"] < PRUNE_TH * 3].index, inplace=True)
print(len(slots), "SLOTS FOUND")
plot_images()
return slots
def look_for_slots(data,
img=[],
PRUNE_TH=3,
ASSIGN_TH=0.6,
plot=True,
PRUNE_STEP=10,
MERGE_STEP=20,
MERGE_TH=0.75):
n_fr = data["frame"].nunique()
cols = [
"labels", 'x1', 'y1', 'x2', 'y2', 'xc', 'yc', 'w', 'b', 'a', "class"
]
base_col = ['x1', 'y1', 'x2', 'y2', 'xc', 'yc', 'w', 'b', 'a']
slots = data[data["frame"] == 0][cols]
slots["found"] = 1
# out_boxes, out_classes, found, labels
# "empty":"#4a148c","occupy":"#f44336", "new":"#7cb342","del":"#80deea"
print("LOOKING FOR PARKING SLOTS INSIDE IMAGE FRAMES")
for i in tqdm(range(1, n_fr)):
post = data[data["frame"] == i].reset_index(drop=True)
_, iou, id_map, status = assign_next_frame(slots, post, th=ASSIGN_TH)
#print(id_map.keys(), status.sum())
## found again
mask = post["labels"].isin(id_map.keys())
slots.loc[status, "found"] = slots.loc[status, "found"] + 1
occupy = post[mask]
occupy["labels"] = occupy["labels"].map(id_map)
slots.sort_values(by=["labels"], inplace=True)
slots.reset_index(drop=True, inplace=True)
occupy.sort_values(by=["labels"], inplace=True)
occupy.reset_index(drop=True, inplace=True)
slots.loc[status, base_col] = slots.loc[status, base_col].values * (
1 - 1 / (i + 1)) + occupy[base_col].values / (i + 1)
# clean up
if i % PRUNE_STEP == 0:
slots.drop(slots[slots["found"] < PRUNE_TH + 1].index,
inplace=True)
#print(slots)
# merge
if i % MERGE_STEP == 0:
slots = compute_distance(slots,
img[i - 1],
th=MERGE_TH,
label="Parking Slots " + str(i))
# new
idx = np.logical_not(post["labels"].isin(id_map.keys()))
new = post[idx]
new["labels"] = new["labels"] + slots["labels"].max() + 1
new = new[cols]
if len(new) > 0:
new["found"] = 1
slots = slots.append(new, sort=True).reset_index(drop=True)
if plot | (i % MERGE_STEP == 0):
df = slots[['x1', 'y1', 'x2', 'y2', "found", "labels"]].copy()
# colors = [(0,1,0)]*len(df)
msk = df["labels"].isin(id_map.values())
colors = [(1, 0, 0) if ms else (0, 1, 0) for ms in msk]
# df.loc[msk,["R","G","B"]] =np.array([(1,0,0)]*msk.sum())
nw_df = new[['x1', 'y1', 'x2', 'y2', "labels"]].copy()
colors = colors + [(0, 1, 0)] * len(nw_df)
nw_df["found"] = 1
df = df.append(nw_df, sort=True).reset_index(drop=True)
df["captions"] = df["labels"].astype(
str) + " [" + df["found"].astype(str) + "]"
# pdb.set_trace()
masks = np.empty((4, 4, len(df)))
class_ids = np.array([3] * len(df))
captions = df["captions"].tolist()
visualize.display_instances(cv2.imread(img[i]),
df[["y1","x1","y2","x2"]].values,\
masks, class_ids, None,
scores=None, title=img[i],
figsize=(20, 20), ax=None,
show_mask=False, show_bbox=True,
colors=colors,
captions=captions)
slots.drop(slots[slots["found"] < PRUNE_TH * 3].index, inplace=True)
slots = compute_distance(slots,
img[0],
th=MERGE_TH * 0.8,
label="Parking Slots " + str(MERGE_STEP))
print(len(slots), "SLOTS FOUND")
return slots