def test_compute_intersect_2d(self):
import time
from det3.ops import compute_intersect_2d, write_pkl
from det3.utils.utils import compute_intersec
for test_dtype in [np.float32, np.float64]:
# npy
boxes = np.array([[3, 2, 4, 4.5]])
boxes = [boxes + np.abs(np.random.randn(4).reshape(1, 4)) for i in range(10)]
boxes = np.vstack(boxes).astype(test_dtype)
others = np.array([[6, 3, 2, 5]])
others = [others + np.abs(np.random.randn(4).reshape(1, 4)) for i in range(21)]
others = np.vstack(others).astype(test_dtype)
boxes_gt = [np.array([itm[0], itm[1], 0, itm[2], itm[3], 0, 0]) for itm in boxes]
boxes_gt = np.vstack(boxes_gt)
others_gt = [np.array([itm[0], itm[1], 0, itm[2], itm[3], 0, 0]) for itm in others]
others_gt = np.vstack(others_gt)
gt = []
for box_gt in boxes_gt:
gt.append(compute_intersec(box_gt, others_gt, mode="2d-rot"))
gt = np.stack(gt, axis=0)
times = 100
est = compute_intersect_2d(boxes, others)
self.assertTrue(est.shape == gt.shape)
self.assertTrue(np.allclose(est, gt, atol=1e-1))
self.assertTrue(est.dtype == boxes.dtype)
t1 = time.time()
for i in range(times):
est = compute_intersect_2d(boxes, others)
t = (time.time()-t1) / times * 1000
print(f"compute_intersect_2d np: {t:.2f} ms")
# torch
boxes_ts = torch.from_numpy(boxes)
others_ts = torch.from_numpy(others)
est_ts = compute_intersect_2d(boxes_ts, others_ts)
self.assertTrue(est_ts.numpy().shape == gt.shape)
self.assertTrue(np.allclose(est_ts.numpy(), gt, atol=1e-1))
self.assertTrue(est_ts.dtype == boxes_ts.dtype)
t1 = time.time()
for i in range(times):
est_ts = compute_intersect_2d(boxes_ts, others_ts)
t = (time.time()-t1) / times * 1000
print(f"compute_intersect_2d torch: {t:.2f} ms")
# torch cuda
boxes_tsgpu = boxes_ts.cuda()
others_tsgpu = others_ts.cuda()
est_tsgpu = compute_intersect_2d(boxes_tsgpu, others_tsgpu)
self.assertTrue(est_tsgpu.cpu().numpy().shape == gt.shape)
self.assertTrue(np.allclose(est_tsgpu.cpu().numpy(), gt, atol=1e-1))
self.assertTrue(est_tsgpu.dtype == boxes_tsgpu.dtype)
torch.cuda.synchronize()
t1 = time.time()
for i in range(times):
est_tsgpu = compute_intersect_2d(boxes_tsgpu, others_tsgpu)
torch.cuda.synchronize()
t = (time.time()-t1) / times * 1000
print(f"compute_intersect_2d torch cuda: {t:.2f} ms")
def test_compute_intersect_3drot(self):
import time
from det3.ops import compute_intersect_3drot, write_pkl
from det3.utils.utils import compute_intersec
for test_dtype in [np.float32, np.float64]:
# npy
boxes = np.array([[5, 3.5, 1, 3, 5, 2, 0.05]])
boxes = [boxes + np.abs(np.random.randn(7).reshape(-1, 7)) for i in range(10)]
boxes = np.vstack(boxes).astype(test_dtype)
others = np.array([[6, 3, 1, 2, 5, 2, 0.05]])
others = [others + np.abs(np.random.randn(7).reshape(-1, 7)) for i in range(21)]
others = np.vstack(others).astype(test_dtype)
gt = []
for box_gt in boxes:
gt.append(compute_intersec(box_gt, others, mode="3d-rot"))
gt = np.stack(gt, axis=0)
times = 100
est = compute_intersect_3drot(boxes, others)
self.assertTrue(est.shape == gt.shape)
self.assertTrue(np.allclose(est, gt, atol=1e-1))
self.assertTrue(est.dtype == boxes.dtype)
t1 = time.time()
for i in range(times):
est = compute_intersect_3drot(boxes, others)
t = (time.time()-t1) / times * 1000
print(f"compute_intersect_3drot np: {t:.2f} ms")
# torch
boxes_ts = torch.from_numpy(boxes)
others_ts = torch.from_numpy(others)
est_ts = compute_intersect_3drot(boxes_ts, others_ts)
self.assertTrue(est_ts.numpy().shape == gt.shape)
self.assertTrue(np.allclose(est_ts.numpy(), gt, atol=1e-1))
self.assertTrue(est_ts.dtype == boxes_ts.dtype)
t1 = time.time()
for i in range(times):
est_ts = compute_intersect_3drot(boxes_ts, others_ts)
t = (time.time()-t1) / times * 1000
print(f"compute_intersect_3drot torch: {t:.2f} ms")
# # torch cuda
boxes_tsgpu = torch.from_numpy(boxes).cuda()
others_tsgpu = torch.from_numpy(others).cuda()
est_tsgpu = compute_intersect_3drot(boxes_tsgpu, others_tsgpu)
self.assertTrue(est_tsgpu.cpu().numpy().shape == gt.shape)
self.assertTrue(np.allclose(est_tsgpu.cpu().numpy(), gt, atol=1e-1))
self.assertTrue(est_tsgpu.dtype == boxes_tsgpu.dtype)
torch.cuda.synchronize()
t1 = time.time()
for i in range(times):
est_tsgpu = compute_intersect_3drot(boxes_tsgpu, others_tsgpu)
torch.cuda.synchronize()
t = (time.time()-t1) / times * 1000
print(f"compute_intersect_3drot torch cuda: {t:.2f} ms")
def check_overlap(label: KittiLabel) -> bool:
'''
check if there is overlap in label.
inputs:
label: the result after agmentation
return:
bool: True if no overlap exists.
'''
assert istype(label, "KittiLabel")
boxes = []
for obj in label.data:
boxes.append([obj.x, obj.z, -obj.y, obj.l, obj.w, obj.h,
-obj.ry]) # Fcam to Flidar
boxes = np.vstack(boxes)
while boxes.shape[0] > 0:
box = boxes[0:1, :]
others = boxes[1:, :]
inter = compute_intersec(box, others, mode='2d-rot')
if inter.sum() > 0:
return False
boxes = others
return True
import torch
import iou_cpp
from det3.utils.utils import compute_intersec
if __name__ == "__main__":
import numpy as np
box = np.array([3, 2, 4, 4.5])
others = np.array([[6, 3, 2, 5], [5, 3, 2, 5]])
others = [others + np.abs(np.random.randn(8).reshape(2, 4)) for i in range(2)]
others = np.vstack(others)
box_ts = torch.from_numpy(box).float()
others_ts = torch.from_numpy(others).float()
est_ts = iou_cpp.compute_intersect_2d_cpu(box_ts, others_ts)
box_gt = np.array([box[0], box[1], 0, box[2], box[3], 0, 0])
others_gt = [np.array([itm[0], itm[1], 0, itm[2], itm[3], 0, 0]) for itm in others]
others_gt = np.vstack(others_gt)
gt = compute_intersec(box_gt, others_gt, mode="2d-rot")
print(est_ts, gt)