File size: 2,380 Bytes
d083e40
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
"""Component-level CHM gating on saved synthetic arrays."""
import cv2
import numpy as np
import torch
import torch.nn.functional as F
import sys
sys.argv = sys.argv[:1]
from chm_eval_score import score  # noqa: E402

d = np.load("/workspace/eval_arr.npz")
pt, lt, nd = d["pt"].astype(np.float32), d["lt"], d["nd"]


def sm(h, k=5):
    return F.avg_pool2d(torch.from_numpy(h.astype(np.float32))[:, None], k, 1, k // 2, count_include_pad=False)[:, 0].numpy()


def gate(pt, hp, hq, drop_min, pre_min, add=None):
    out = []
    for p, a, b in zip(pt, hp, hq):
        m = (p > 0.5).astype(np.uint8)
        n, cc, st, _ = cv2.connectedComponentsWithStats(m, 8)
        keep = m.astype(bool).copy()
        for i in range(1, n):
            r = cc == i
            if (a[r] - b[r]).mean() < drop_min and a[r].mean() < pre_min:
                keep[r] = False
        if add is not None:   # add strong CHM removal blobs where the ensemble is at least weakly positive
            hmin, gmax, pmin, amin = add
            c = ((a > hmin) & (b < gmax) & (p > pmin)).astype(np.uint8)
            n2, cc2, st2, _ = cv2.connectedComponentsWithStats(c, 8)
            for i in range(1, n2):
                if st2[i, cv2.CC_STAT_AREA] >= amin:
                    keep |= cc2 == i
        out.append(keep.astype(np.float32))
    return np.stack(out)


for tag, hp, hq in (("x1.0", d["hp"], d["hq"]), ("x1.5", d["hp2"], d["hq2"])):
    hp, hq = sm(hp), sm(hq)
    print(tag, "base", round(score(pt, lt)[0], 4), flush=True)
    for dm, pm in ((0.5, 2.0), (1.0, 2.5), (1.5, 3.0), (2.0, 4.0)):
        s, c = score(gate(pt, hp, hq, dm, pm), lt)
        print(f"{tag} drop-gate drop<{dm} & pre<{pm}: {s:.4f} {c}", flush=True)
    for add in ((5, 2, 0.3, 40), (5, 2, 0.2, 60), (6, 1.5, 0.15, 80), (4, 2, 0.35, 30)):
        s, c = score(gate(pt, hp, hq, 1.0, 2.5, add), lt)
        print(f"{tag} drop-gate(1.0,2.5)+add{add}: {s:.4f} {c}", flush=True)
    # how many GT-negative pairs would a pure CHM blob detector flag
    fp = sum(((a > 5) & (b < 2)).sum() >= 60 for a, b, g in zip(hp, hq, lt) if g.sum() < 20)
    tp = sum(((a > 5) & (b < 2)).sum() >= 60 for a, b, g in zip(hp, hq, lt) if g.sum() >= 20)
    print(f"{tag} CHM-only blob(>5m -> <2m, >=60px): flags {tp}/{int(sum(g.sum() >= 20 for g in lt))} pos, {fp}/{int(sum(g.sum() < 20 for g in lt))} neg", flush=True)
print("GATE_DONE")