Spaces:
Running
Running
Added better visual
Browse files- app.py +7 -13
- ppe_compliance.py +69 -78
app.py
CHANGED
|
@@ -157,11 +157,11 @@ with gr.Blocks(title="Small Object Detection") as app:
|
|
| 157 |
with gr.Tab("PPE Compliance"):
|
| 158 |
|
| 159 |
gr.Markdown(
|
| 160 |
-
"A **D-FINE person detector** (same model as the Detect & Classify tab)
|
| 161 |
-
"**fine-tuned D-FINE-M PPE detector**
|
| 162 |
-
"
|
| 163 |
-
"
|
| 164 |
-
"
|
| 165 |
)
|
| 166 |
|
| 167 |
with gr.Row():
|
|
@@ -190,11 +190,6 @@ with gr.Blocks(title="Small Object Detection") as app:
|
|
| 190 |
label="PPE detection threshold",
|
| 191 |
)
|
| 192 |
|
| 193 |
-
region_checkbox = gr.Checkbox(
|
| 194 |
-
value=True,
|
| 195 |
-
label="Body-region check (only count PPE worn in the expected region)",
|
| 196 |
-
)
|
| 197 |
-
|
| 198 |
btn_ppe = gr.Button(
|
| 199 |
"Run PPE Compliance",
|
| 200 |
variant="primary",
|
|
@@ -214,13 +209,12 @@ with gr.Blocks(title="Small Object Detection") as app:
|
|
| 214 |
)
|
| 215 |
|
| 216 |
btn_ppe.click(
|
| 217 |
-
fn=lambda image, p_thr, ppe_thr
|
| 218 |
image,
|
| 219 |
person_threshold=float(p_thr),
|
| 220 |
ppe_threshold=float(ppe_thr),
|
| 221 |
-
use_region=bool(region),
|
| 222 |
),
|
| 223 |
-
inputs=[inp_ppe, person_threshold_slider, ppe_threshold_slider
|
| 224 |
outputs=[out_ppe_image, out_ppe_status],
|
| 225 |
concurrency_limit=1,
|
| 226 |
)
|
|
|
|
| 157 |
with gr.Tab("PPE Compliance"):
|
| 158 |
|
| 159 |
gr.Markdown(
|
| 160 |
+
"A **D-FINE person detector** (same model as the Detect & Classify tab) finds each "
|
| 161 |
+
"person, then the **fine-tuned D-FINE-M PPE detector** runs on each person crop. "
|
| 162 |
+
"Any of the 6 required items (goggles, helmet, mask, shoes, vest, glove) that are "
|
| 163 |
+
"missing for a person are flagged. **Green** box = compliant, **red** box = violation; "
|
| 164 |
+
"a per-person checklist (1 = present, 0 = missing) is drawn beside each box."
|
| 165 |
)
|
| 166 |
|
| 167 |
with gr.Row():
|
|
|
|
| 190 |
label="PPE detection threshold",
|
| 191 |
)
|
| 192 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 193 |
btn_ppe = gr.Button(
|
| 194 |
"Run PPE Compliance",
|
| 195 |
variant="primary",
|
|
|
|
| 209 |
)
|
| 210 |
|
| 211 |
btn_ppe.click(
|
| 212 |
+
fn=lambda image, p_thr, ppe_thr: run_ppe_compliance(
|
| 213 |
image,
|
| 214 |
person_threshold=float(p_thr),
|
| 215 |
ppe_threshold=float(ppe_thr),
|
|
|
|
| 216 |
),
|
| 217 |
+
inputs=[inp_ppe, person_threshold_slider, ppe_threshold_slider],
|
| 218 |
outputs=[out_ppe_image, out_ppe_status],
|
| 219 |
concurrency_limit=1,
|
| 220 |
)
|
ppe_compliance.py
CHANGED
|
@@ -13,12 +13,13 @@ items are MISSING (the "negation" logic):
|
|
| 13 |
like vests), we run it on EACH person crop and map the boxes back. On this
|
| 14 |
image that lifted vest detections from 0 -> 4.
|
| 15 |
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
|
|
|
| 22 |
"""
|
| 23 |
|
| 24 |
import torch
|
|
@@ -33,18 +34,6 @@ DEVICE = "cpu"
|
|
| 33 |
|
| 34 |
ALL_PPE = ["goggles", "helmet", "mask", "shoes", "vest", "glove"]
|
| 35 |
|
| 36 |
-
# Vertical band (fraction of person-box height from the top) where each PPE item
|
| 37 |
-
# is expected. A detection only counts for a person if its center falls in the
|
| 38 |
-
# band β this rejects e.g. a helmet a worker is *holding* (low) vs *wearing* (top).
|
| 39 |
-
REGION_BANDS = {
|
| 40 |
-
"helmet": (0.00, 0.35),
|
| 41 |
-
"goggles": (0.00, 0.40),
|
| 42 |
-
"mask": (0.05, 0.45),
|
| 43 |
-
"vest": (0.20, 0.75),
|
| 44 |
-
"glove": (0.30, 0.95),
|
| 45 |
-
"shoes": (0.70, 1.00),
|
| 46 |
-
}
|
| 47 |
-
|
| 48 |
# Person detector defaults to the same model the "Detect & Classify" tab uses.
|
| 49 |
DEFAULT_PERSON_MODEL = "medium-obj2coco"
|
| 50 |
|
|
@@ -132,7 +121,6 @@ def run_ppe_compliance(
|
|
| 132 |
person_threshold=0.5,
|
| 133 |
ppe_threshold=0.4,
|
| 134 |
assoc=0.5,
|
| 135 |
-
use_region=True,
|
| 136 |
person_model=DEFAULT_PERSON_MODEL,
|
| 137 |
min_side=960,
|
| 138 |
):
|
|
@@ -160,11 +148,17 @@ def run_ppe_compliance(
|
|
| 160 |
persons.sort(key=lambda ps: -ps[0])
|
| 161 |
persons = persons[:MAX_PERSON_CROPS]
|
| 162 |
|
| 163 |
-
# 2) PPE detector (fine-tuned D-FINE-M), run on EACH person crop so small PPE
|
| 164 |
-
# (vests, distant helmets) isn't lost to the 640x640 full-frame resize
|
| 165 |
-
#
|
| 166 |
-
|
| 167 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 168 |
x1, y1, x2, y2 = pb
|
| 169 |
pw, ph = x2 - x1, y2 - y1
|
| 170 |
cx1 = max(0, int(x1 - PERSON_CROP_PAD * pw))
|
|
@@ -175,29 +169,15 @@ def run_ppe_compliance(
|
|
| 175 |
continue
|
| 176 |
crop = im.crop((cx1, cy1, cx2, cy2))
|
| 177 |
for name, s, b in detect_ppe_boxes(crop, threshold=ppe_threshold):
|
| 178 |
-
|
| 179 |
-
|
| 180 |
-
|
| 181 |
-
|
| 182 |
-
people = [{"score": s, "box": box, "present": {}} for s, box in persons]
|
| 183 |
-
for name, s, box in ppe:
|
| 184 |
-
cy = 0.5 * (box[1] + box[3]) # PPE box vertical center
|
| 185 |
-
best_i, best_f = -1, assoc
|
| 186 |
-
for i, p in enumerate(people):
|
| 187 |
-
f = _contain_frac(box, p["box"])
|
| 188 |
-
if f < best_f:
|
| 189 |
continue
|
| 190 |
-
|
| 191 |
-
py1, py2 = p["box"][1], p["box"][3]
|
| 192 |
-
frac = (cy - py1) / max(1e-6, py2 - py1) # where in the person it sits
|
| 193 |
-
lo, hi = REGION_BANDS[name]
|
| 194 |
-
if not (lo <= frac <= hi):
|
| 195 |
-
continue # right person, wrong body region -> not "worn"
|
| 196 |
-
best_i, best_f = i, f
|
| 197 |
-
if best_i >= 0:
|
| 198 |
-
cur = people[best_i]["present"].get(name)
|
| 199 |
if cur is None or s > cur[0]:
|
| 200 |
-
people[
|
|
|
|
| 201 |
|
| 202 |
# 4) Verdicts + report.
|
| 203 |
verdicts = []
|
|
@@ -221,49 +201,60 @@ def run_ppe_compliance(
|
|
| 221 |
p["box"] = [v * scale for v in p["box"]]
|
| 222 |
ppe = [(n, s, [v * scale for v in b]) for (n, s, b) in ppe]
|
| 223 |
|
| 224 |
-
# 5) Render:
|
|
|
|
|
|
|
|
|
|
| 225 |
W, H = im.size
|
| 226 |
base = im.convert("RGBA")
|
| 227 |
ov = Image.new("RGBA", (W, H), (0, 0, 0, 0))
|
| 228 |
od = ImageDraw.Draw(ov)
|
| 229 |
-
|
| 230 |
-
|
| 231 |
-
placed = [] # occupied
|
| 232 |
-
|
| 233 |
-
def
|
| 234 |
-
return
|
| 235 |
-
|
| 236 |
-
def chip(x, y, text, font, bg, above=True):
|
| 237 |
-
tb = od.textbbox((0, 0), text, font=font)
|
| 238 |
-
tw, th = tb[2] - tb[0], tb[3] - tb[1]
|
| 239 |
-
pad = 4
|
| 240 |
-
rx = int(min(max(0, x), max(0, W - tw - 2 * pad)))
|
| 241 |
-
ry = max(0, int(y - th - 2 * pad) if above else int(y))
|
| 242 |
-
for _ in range(80):
|
| 243 |
-
r = (rx, ry, rx + tw + 2 * pad, ry + th + 2 * pad)
|
| 244 |
-
if fits(r) and r[3] <= H:
|
| 245 |
-
break
|
| 246 |
-
ry += th + 2 * pad + 2
|
| 247 |
-
r = (rx, ry, rx + tw + 2 * pad, ry + th + 2 * pad)
|
| 248 |
-
od.rounded_rectangle(r, radius=5, fill=bg + (230,))
|
| 249 |
-
od.text((rx + pad, ry + pad - 1), text, font=font, fill=(255, 255, 255, 255))
|
| 250 |
-
placed.append(r)
|
| 251 |
|
| 252 |
-
#
|
| 253 |
for name, s, box in ppe:
|
| 254 |
od.rectangle(box, outline=_color_of(name) + (255,), width=2)
|
| 255 |
|
| 256 |
-
# 2) thick person boxes + verdict chips (placed first = priority)
|
| 257 |
pw = max(3, int(0.006 * max(W, H)))
|
|
|
|
|
|
|
| 258 |
for i, p, present, missing, ok in verdicts:
|
| 259 |
-
|
| 260 |
-
x1, y1, x2, y2 = p["box"]
|
| 261 |
-
od.rectangle([x1, y1, x2, y2], outline=
|
| 262 |
-
|
| 263 |
-
|
| 264 |
-
|
| 265 |
-
|
| 266 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 267 |
|
| 268 |
out_im = Image.alpha_composite(base, ov).convert("RGB")
|
| 269 |
return out_im, "\n".join(lines)
|
|
|
|
| 13 |
like vests), we run it on EACH person crop and map the boxes back. On this
|
| 14 |
image that lifted vest detections from 0 -> 4.
|
| 15 |
|
| 16 |
+
Each PPE detection is credited to the person whose crop produced it (gated by
|
| 17 |
+
containment in that person's box). A person missing any required item is a
|
| 18 |
+
VIOLATION. Rendering: green box = OK, red box = violation, with a compact
|
| 19 |
+
per-person checklist (1=present / 0=missing) drawn outside the box. Adapted from
|
| 20 |
+
``ppe/scripts/compliance.py``; the original's body-region bands are dropped β
|
| 21 |
+
per-person cropping already localises PPE to the right worker, so the band only
|
| 22 |
+
hurt recall on bent/crouched poses.
|
| 23 |
"""
|
| 24 |
|
| 25 |
import torch
|
|
|
|
| 34 |
|
| 35 |
ALL_PPE = ["goggles", "helmet", "mask", "shoes", "vest", "glove"]
|
| 36 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
# Person detector defaults to the same model the "Detect & Classify" tab uses.
|
| 38 |
DEFAULT_PERSON_MODEL = "medium-obj2coco"
|
| 39 |
|
|
|
|
| 121 |
person_threshold=0.5,
|
| 122 |
ppe_threshold=0.4,
|
| 123 |
assoc=0.5,
|
|
|
|
| 124 |
person_model=DEFAULT_PERSON_MODEL,
|
| 125 |
min_side=960,
|
| 126 |
):
|
|
|
|
| 148 |
persons.sort(key=lambda ps: -ps[0])
|
| 149 |
persons = persons[:MAX_PERSON_CROPS]
|
| 150 |
|
| 151 |
+
# 2+3) PPE detector (fine-tuned D-FINE-M), run on EACH person crop so small PPE
|
| 152 |
+
# (vests, distant helmets) isn't lost to the 640x640 full-frame resize, then
|
| 153 |
+
# attribute the detections to THAT person. We credit a detection to the person
|
| 154 |
+
# whose crop produced it (gated by containment in their own box),
|
| 155 |
+
# rather than re-assigning globally to whichever overlapping neighbour contains
|
| 156 |
+
# it most β that global step mislabelled a worker's own vest to a denser
|
| 157 |
+
# neighbour in crowded frames, so the vest was drawn but the wearer showed
|
| 158 |
+
# "no vest". Boxes are mapped back to full coords; the draw list is de-duped.
|
| 159 |
+
people = [{"score": conf, "box": pb, "present": {}} for conf, pb in persons]
|
| 160 |
+
ppe = [] # (name, score, full_box) β for drawing only
|
| 161 |
+
for i, (conf, pb) in enumerate(persons):
|
| 162 |
x1, y1, x2, y2 = pb
|
| 163 |
pw, ph = x2 - x1, y2 - y1
|
| 164 |
cx1 = max(0, int(x1 - PERSON_CROP_PAD * pw))
|
|
|
|
| 169 |
continue
|
| 170 |
crop = im.crop((cx1, cy1, cx2, cy2))
|
| 171 |
for name, s, b in detect_ppe_boxes(crop, threshold=ppe_threshold):
|
| 172 |
+
box = [b[0] + cx1, b[1] + cy1, b[2] + cx1, b[3] + cy1]
|
| 173 |
+
ppe.append((name, s, box))
|
| 174 |
+
# Attribute to this crop's person: must lie mostly inside their box.
|
| 175 |
+
if _contain_frac(box, pb) < assoc:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 176 |
continue
|
| 177 |
+
cur = people[i]["present"].get(name)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 178 |
if cur is None or s > cur[0]:
|
| 179 |
+
people[i]["present"][name] = (s, box)
|
| 180 |
+
ppe = _dedup_ppe(ppe)
|
| 181 |
|
| 182 |
# 4) Verdicts + report.
|
| 183 |
verdicts = []
|
|
|
|
| 201 |
p["box"] = [v * scale for v in p["box"]]
|
| 202 |
ppe = [(n, s, [v * scale for v in b]) for (n, s, b) in ppe]
|
| 203 |
|
| 204 |
+
# 5) Render: thin PPE boxes + a compact per-person CHECKLIST panel placed
|
| 205 |
+
# OUTSIDE the person box. One short row per required item ("vest 1" present /
|
| 206 |
+
# "vest 0" missing, green/red), stacked in a column β instead of one long
|
| 207 |
+
# "no goggles, mask, shoes, glove" sentence that collides in crowded frames.
|
| 208 |
W, H = im.size
|
| 209 |
base = im.convert("RGBA")
|
| 210 |
ov = Image.new("RGBA", (W, H), (0, 0, 0, 0))
|
| 211 |
od = ImageDraw.Draw(ov)
|
| 212 |
+
rf = _load_font(max(12, int(0.016 * max(W, H)))) # checklist rows
|
| 213 |
+
GREEN, RED = (40, 200, 90), (235, 45, 55)
|
| 214 |
+
placed = [] # occupied panel rects, to avoid overlaps
|
| 215 |
+
|
| 216 |
+
def overlaps(r):
|
| 217 |
+
return any(not (r[2] <= q[0] or r[0] >= q[2] or r[3] <= q[1] or r[1] >= q[3]) for q in placed)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 218 |
|
| 219 |
+
# thin PPE boxes (no inline text β presence is shown in the checklist)
|
| 220 |
for name, s, box in ppe:
|
| 221 |
od.rectangle(box, outline=_color_of(name) + (255,), width=2)
|
| 222 |
|
|
|
|
| 223 |
pw = max(3, int(0.006 * max(W, H)))
|
| 224 |
+
pad = 5
|
| 225 |
+
line_h = rf.getbbox("Hg1")[3] + 4
|
| 226 |
for i, p, present, missing, ok in verdicts:
|
| 227 |
+
col = GREEN if ok else RED
|
| 228 |
+
x1, y1, x2, y2 = [int(v) for v in p["box"]]
|
| 229 |
+
od.rectangle([x1, y1, x2, y2], outline=col + (255,), width=pw)
|
| 230 |
+
|
| 231 |
+
rows = [(f"#{i}", None)] + [(it, it in present) for it in ALL_PPE]
|
| 232 |
+
txts = [lab if pres is None else f"{lab} {1 if pres else 0}" for lab, pres in rows]
|
| 233 |
+
p_w = max(od.textbbox((0, 0), t, font=rf)[2] for t in txts) + 2 * pad
|
| 234 |
+
p_h = line_h * len(rows) + 2 * pad
|
| 235 |
+
|
| 236 |
+
# Prefer just right of the box; fall back to the left; keep inside frame.
|
| 237 |
+
rx = x2 + 6
|
| 238 |
+
if rx + p_w > W:
|
| 239 |
+
rx = x1 - p_w - 6
|
| 240 |
+
if rx < 0:
|
| 241 |
+
rx = min(x1 + 6, max(0, W - p_w))
|
| 242 |
+
ry = max(0, min(y1, H - p_h))
|
| 243 |
+
for _ in range(300): # nudge down, then to a fresh column, until it fits
|
| 244 |
+
if not overlaps((rx, ry, rx + p_w, ry + p_h)):
|
| 245 |
+
break
|
| 246 |
+
ry += line_h
|
| 247 |
+
if ry + p_h > H:
|
| 248 |
+
ry = max(0, min(y1, H - p_h))
|
| 249 |
+
rx = rx + p_w + 6
|
| 250 |
+
if rx + p_w > W:
|
| 251 |
+
rx = max(0, W - p_w)
|
| 252 |
+
placed.append((rx, ry, rx + p_w, ry + p_h))
|
| 253 |
+
|
| 254 |
+
od.rounded_rectangle((rx, ry, rx + p_w, ry + p_h), radius=5, fill=(0, 0, 0, 175))
|
| 255 |
+
for k, (lab, pres) in enumerate(rows):
|
| 256 |
+
tc = col if pres is None else (GREEN if pres else RED)
|
| 257 |
+
od.text((rx + pad, ry + pad + k * line_h), txts[k], font=rf, fill=tc + (255,))
|
| 258 |
|
| 259 |
out_im = Image.alpha_composite(base, ov).convert("RGB")
|
| 260 |
return out_im, "\n".join(lines)
|