def demo(): net = Resnet18_8s(ver_dim=vote_num * 2, seg_dim=2) net = NetWrapper(net).cuda() net = DataParallel(net) optimizer = optim.Adam(net.parameters(), lr=train_cfg['lr']) model_dir = os.path.join(cfg.MODEL_DIR, "cat_demo") load_model(net.module.net, optimizer, model_dir, args.load_epoch) data, points_3d, bb8_3d = read_data() image, mask, vertex, vertex_weights, pose, corner_target = [ d.unsqueeze(0).cuda() for d in data ] seg_pred, vertex_pred, loss_seg, loss_vertex, precision, recall = net( image, mask, vertex, vertex_weights) eval_net = DataParallel(EvalWrapper().cuda()) corner_pred = eval_net(seg_pred, vertex_pred).cpu().detach().numpy()[0] camera_matrix = np.array([[572.4114, 0., 325.2611], [0., 573.57043, 242.04899], [0., 0., 1.]]) pose_pred = pnp(points_3d, corner_pred, camera_matrix) projector = Projector() bb8_2d_pred = projector.project(bb8_3d, pose_pred, 'linemod') bb8_2d_gt = projector.project(bb8_3d, pose[0].detach().cpu().numpy(), 'linemod') image = imagenet_to_uint8(image.detach().cpu().numpy())[0] visualize_bounding_box(image[None, ...], bb8_2d_pred[None, None, ...], bb8_2d_gt[None, None, ...])
def demo(): net = Resnet18_8s(ver_dim=vote_num * 2, seg_dim=2) net = NetWrapper(net).cuda() net = DataParallel(net) optimizer = optim.Adam(net.parameters(), lr=train_cfg['lr']) model_dir = os.path.join(cfg.MODEL_DIR, "switch_linemod_train") load_model(net.module.net, optimizer, model_dir, -1) image, points_3d, bb8_3d = read_data() image = image[None, ...] seg_pred, vertex_pred = net(image) # visualize_mask(mask) # visualize_vertex(vertex, vertex_weights) # visualize_hypothesis(image, seg_pred, vertex_pred, corner_target) # visualize_voting_ellipse(image, seg_pred, vertex_pred, corner_target) eval_net = DataParallel(EvalWrapper().cuda()) corner_pred = eval_net(seg_pred, vertex_pred).cpu().detach().numpy()[0] camera_matrix = np.array([[572.4114, 0., 325.2611], [0., 573.57043, 242.04899], [0., 0., 1.]]) pose_pred = pnp(points_3d, corner_pred, camera_matrix) projector = Projector() bb8_2d_pred = projector.project(bb8_3d, pose_pred, 'linemod') print(bb8_2d_pred) image = imagenet_to_uint8(image.detach().cpu().numpy())[0] visualize_bounding_box(image[None, ...], bb8_2d_pred[None, None, ...])
def demo(): net = Resnet18_8s(ver_dim=vote_num * 2, seg_dim=2) net = NetWrapper(net).cuda() net = DataParallel(net) optimizer = optim.Adam(net.parameters(), lr=train_cfg['lr']) model_dir = os.path.join(cfg.MODEL_DIR, "cat_linemod_train") #cat_demo load_model(net.module.net, optimizer, model_dir, args.load_epoch) data, points_3d, bb8_3d = read_data() image, mask, vertex, vertex_weights, pose, corner_target = [ d.unsqueeze(0).cuda() for d in data ] seg_pred, vertex_pred, loss_seg, loss_vertex, precision, recall = net( image, mask, vertex, vertex_weights) eval_net = DataParallel(EvalWrapper().cuda()) #向量方形图,语义分割图,然后 ransac 计算 kp,,向量方向图一旦准了,kp也就准了 corner_pred = eval_net(seg_pred, vertex_pred).cpu().detach().numpy()[0] camera_matrix = np.array([[572.4114, 0., 325.2611], [0., 573.57043, 242.04899], [0., 0., 1.]]) pose_pred = pnp(points_3d, corner_pred, camera_matrix) projector = Projector() # bb8_2d_pred = projector.project(bb8_3d, pose_pred, 'linemod') bb8_2d_gt = projector.project(bb8_3d, pose[0].detach().cpu().numpy(), 'linemod') image = imagenet_to_uint8(image.detach().cpu().numpy())[0] print("loss_seg:{} , loss_vertex:{} , precision:{},recall:{}, ".format( loss_seg, loss_vertex, precision, recall)) #399.pth #loss_seg:tensor([0.0015], device='cuda:0', grad_fn=<MeanBackward0>) , loss_vertex:tensor([0.0016], device='cuda:0', grad_fn=<DivBackward1>) , #precision:tensor([0.9434], device='cuda:0'),recall:tensor([0.9677], device='cuda:0'), #199.pth # loss_seg:tensor([0.0015], device='cuda:0', grad_fn=<MeanBackward0>) , loss_vertex:tensor([0.0016], device='cuda:0', grad_fn=<DivBackward1>) , # precision:tensor([0.9583], device='cuda:0'),recall:tensor([0.9524], device='cuda:0'), erro = bb8_2d_pred - bb8_2d_gt erro = np.abs(erro) err = np.reshape(erro, (erro.size, )) #abserr = map(abs,err) print("reproject sum_error:{} ".format(np.sum(err))) ## 199 reproject sum_error:13.385891544820552 ## 399 reproject sum_error:12.718721049803733 ##看了是有提高 准召定义 准确下降,召回上升 visualize_bounding_box(image[None, ...], bb8_2d_pred[None, None, ...], bb8_2d_gt[None, None, ...])
def demo(idx): data, bb8_3d = read_data(idx) print("BB8_3D: ",bb8_3d) image, mask, pose = [d.unsqueeze(0).cuda() for d in data] projector = Projector() bb8_2d_gt = projector.project(bb8_3d, pose[0].detach().cpu().numpy(), 'blender') print(bb8_2d_gt) image = imagenet_to_uint8(image.detach().cpu().numpy())[0] visualize_bounding_box(image[None, ...], bb8_2d_gt[None, None, ...])
def inference(input_image, count=0): c_timer = time.time() rgb = input_image if args.input != 'image': color = cv2.cvtColor(input_image, cv2.COLOR_BGR2RGB) rgb = color pre_start = time.time() print(pre_start - c_timer, "s BGR2RGB") #rgb = Image.open(input_image) #print(rgb.shape) start = time.time() transformer = transforms.Compose([ transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ]) rgb = transformer(rgb) rgb = rgb.unsqueeze(0).cuda() seg_pred, vertex_pred = net(rgb) eval_net = DataParallel(EvalWrapper().cuda()) corner_pred = eval_net(seg_pred, vertex_pred).cpu().detach().numpy()[0] end = time.time() print(end - start, "s - to go from image to corner prediction") image = imagenet_to_uint8(rgb.detach().cpu().numpy())[0] pose_pred = pnp(points_3d, corner_pred, camera_matrix) projector = Projector() bb8_2d_pred = projector.project(bb8_3d, pose_pred, 'logitech') end_ = time.time() print(end_ - end, "s - to project the corners and show the result") seg_mask = torch.argmax(seg_pred, 1) if args.debug: visualize_mask(seg_mask, count) pose_test = np.array([[1, 0, 0, 0], [0, 1, 0, 0.3], [0, 0, 1, 1.2]]) print(pose_pred) #print(pose_test) bb8_2d_gt = projector.project(bb8_3d, pose_test, 'logitech') if pose_pred[2][3] < 0.4: if pose_pred[2][3] > -0.4: if isinstance(rgb, torch.Tensor): rgb = rgb.permute(0, 2, 3, 1).detach().cpu().numpy() rgb = rgb.astype(np.uint8) _, ax = plt.subplots(1) ax.imshow(cv2.cvtColor(input_image, cv2.COLOR_BGR2RGB)) #plt.show() plt.savefig('temp{}.png'.format(count)) plt.close() print("image was culled due to pose being unreasonable") else: visualize_bounding_box(image[None, ...], bb8_2d_pred[None, None, ...], save=True, count=count) #,bb8_2d_gt[None, None, ...])
def visualize_hypothesis(image, seg_pred, vertex_pred, corner_target): from lib.ransac_voting_gpu_layer.ransac_voting_gpu import generate_hypothesis vertex_pred = vertex_pred.permute(0, 2, 3, 1) b, h, w, vn_2 = vertex_pred.shape vertex_pred = vertex_pred.view(b, h, w, vn_2 // 2, 2) mask = torch.argmax(seg_pred, 1) hyp, hyp_counts = generate_hypothesis(mask, vertex_pred, 1024, inlier_thresh=0.99) image = imagenet_to_uint8(image.detach().cpu().numpy()) hyp = hyp.detach().cpu().numpy() hyp_counts = hyp_counts.detach().cpu().numpy() from lib.utils.draw_utils import visualize_hypothesis visualize_hypothesis(image, hyp, hyp_counts, corner_target)
def visualize_voting_ellipse(image, seg_pred, vertex_pred, corner_target): from lib.ransac_voting_gpu_layer.ransac_voting_gpu import estimate_voting_distribution_with_mean vertex_pred = vertex_pred.permute(0, 2, 3, 1) b, h, w, vn_2 = vertex_pred.shape vertex_pred = vertex_pred.view(b, h, w, vn_2 // 2, 2) mask = torch.argmax(seg_pred, 1) mean = ransac_voting_layer_v3(mask, vertex_pred, 512, inlier_thresh=0.99) mean, var = estimate_voting_distribution_with_mean(mask, vertex_pred, mean) image = imagenet_to_uint8(image.detach().cpu().numpy()) mean = mean.detach().cpu().numpy() var = var.detach().cpu().numpy() corner_target = corner_target.detach().cpu().numpy() from lib.utils.draw_utils import visualize_voting_ellipse visualize_voting_ellipse(image, mean, var, corner_target)
def demo(): net = Resnet18_8s(ver_dim=vote_num * 2, seg_dim=2) net = NetWrapper(net).cuda() net = DataParallel(net) optimizer = optim.Adam(net.parameters(), lr=train_cfg['lr']) model_dir = os.path.join(cfg.MODEL_DIR, "cat_demo") load_model(net.module.net, optimizer, model_dir, -1) data, points_3d, bb8_3d = read_data() #print("BB8_3D: ",bb8_3d) image, mask, vertex, vertex_weights, pose, corner_target = [ d.unsqueeze(0).cuda() for d in data ] seg_pred, vertex_pred, loss_seg, loss_vertex, precision, recall = net( image, mask, vertex, vertex_weights) seg_mask = torch.argmax(seg_pred, 1) print("seg_mask", seg_mask, type(seg_mask), seg_mask.shape, seg_mask[0]) visualize_mask(seg_mask) visualize_mask(mask) #visualize_vertex(vertex, vertex_weights) #visualize_hypothesis(image, seg_pred, vertex_pred, corner_target) visualize_voting_ellipse(image, seg_pred, vertex_pred, corner_target) eval_net = DataParallel(EvalWrapper().cuda()) corner_pred = eval_net(seg_pred, vertex_pred).cpu().detach().numpy()[0] print("Corner Predictions: ", corner_pred) camera_matrix = np.array([[572.4114, 0., 325.2611], [0., 573.57043, 242.04899], [0., 0., 1.]]) pose_pred = pnp(points_3d, corner_pred, camera_matrix) projector = Projector() bb8_2d_pred = projector.project(bb8_3d, pose_pred, 'linemod') print("Pose prediction :\n", pose_pred) print("GT pose: \n", pose[0].detach().cpu().numpy()) bb8_2d_gt = projector.project(bb8_3d, pose[0].detach().cpu().numpy(), 'linemod') print(bb8_2d_gt) image = imagenet_to_uint8(image.detach().cpu().numpy())[0] visualize_bounding_box(image[None, ...], bb8_2d_pred[None, None, ...], bb8_2d_gt[None, None, ...])
def val(net, dataloader, epoch, val_prefix='val', use_camera_intrinsic=False, use_motion=False): for rec in recs: rec.reset() test_begin = time.time() evaluator = Evaluator() eval_net = DataParallel( EvalWrapper().cuda()) if not use_motion else DataParallel( MotionEvalWrapper().cuda()) uncertain_eval_net = DataParallel(UncertaintyEvalWrapper().cuda()) net.eval() for idx, data in enumerate(dataloader): if use_camera_intrinsic: image, mask, vertex, vertex_weights, pose, corner_target, Ks = [ d.cuda() for d in data ] else: image, mask, vertex, vertex_weights, pose, corner_target = [ d.cuda() for d in data ] with torch.no_grad(): seg_pred, vertex_pred, loss_seg, loss_vertex, precision, recall = net( image, mask, vertex, vertex_weights) loss_seg, loss_vertex, precision, recall = [ torch.mean(val) for val in (loss_seg, loss_vertex, precision, recall) ] if (train_cfg['eval_epoch'] and epoch % train_cfg['eval_inter'] == 0 and epoch >= train_cfg['eval_epoch_begin']) or args.test_model: if args.use_uncertainty_pnp: mean, cov_inv = uncertain_eval_net(seg_pred, vertex_pred) mean = mean.cpu().numpy() cov_inv = cov_inv.cpu().numpy() else: corner_pred = eval_net(seg_pred, vertex_pred).cpu().detach().numpy() pose = pose.cpu().numpy() b = pose.shape[0] pose_preds = [] for bi in range(b): intri_type = 'use_intrinsic' if use_camera_intrinsic else 'linemod' K = Ks[bi].cpu().numpy() if use_camera_intrinsic else None if args.use_uncertainty_pnp: pose_preds.append( evaluator.evaluate_uncertainty(mean[bi], cov_inv[bi], pose[bi], args.linemod_cls, intri_type, vote_type, intri_matrix=K)) else: pose_preds.append( evaluator.evaluate(corner_pred[bi], pose[bi], args.linemod_cls, intri_type, vote_type, intri_matrix=K)) if args.save_inter_result: mask_pr = torch.argmax(seg_pred, 1).cpu().detach().numpy() mask_gt = mask.cpu().detach().numpy() # assume batch size = 1 imsave( os.path.join(args.save_inter_dir, '{}_mask_pr.png'.format(idx)), mask_pr[0]) imsave( os.path.join(args.save_inter_dir, '{}_mask_gt.png'.format(idx)), mask_gt[0]) imsave( os.path.join(args.save_inter_dir, '{}_rgb.png'.format(idx)), imagenet_to_uint8(image.cpu().detach().numpy()[0])) save_pickle([pose_preds[0], pose[0]], os.path.join(args.save_inter_dir, '{}_pose.pkl'.format(idx))) vals = [loss_seg, loss_vertex, precision, recall] for rec, val in zip(recs, vals): rec.update(val) with torch.no_grad(): batch_size = image.shape[0] nrow = 5 if batch_size > 5 else batch_size recorder.rec_segmentation(F.softmax(seg_pred, dim=1), num_classes=2, nrow=nrow, step=epoch, name='{}/image/seg'.format(val_prefix)) recorder.rec_vertex(vertex_pred, vertex_weights, nrow=4, step=epoch, name='{}/image/ver'.format(val_prefix)) losses_batch = OrderedDict() for name, rec in zip(recs_names, recs): losses_batch['{}/'.format(val_prefix) + name] = rec.avg if (train_cfg['eval_epoch'] and epoch % train_cfg['eval_inter'] == 0 and epoch >= train_cfg['eval_epoch_begin']) or args.test_model: proj_err, add, cm = evaluator.average_precision(False) losses_batch['{}/scalar/projection_error'.format( val_prefix)] = proj_err losses_batch['{}/scalar/add'.format(val_prefix)] = add losses_batch['{}/scalar/cm'.format(val_prefix)] = cm recorder.rec_loss_batch(losses_batch, epoch, epoch, val_prefix) for rec in recs: rec.reset() print('epoch {} {} cost {} s'.format(epoch, val_prefix, time.time() - test_begin))
# camera_matrix = np.array([[572.4114, 0., 325.2611], # [0., 573.57043, 242.04899], # [0., 0., 1.]]) # pose_pred = pnp(points_3d, corner_pred, camera_matrix) # projector = Projector() # bb8_2d_pred = projector.project(bb8_3d, pose_pred, 'linemod') # bb8_2d_gt = projector.project(bb8_3d, pose[0].detach().cpu().numpy(), 'linemod') # image = imagenet_to_uint8(image.detach().cpu().numpy())[0] # visualize_bounding_box(image[None, ...], bb8_2d_pred[None, None, ...], bb8_2d_gt[None, None, ...]) ======= seg_pred, vertex_pred, loss_seg, loss_vertex, precision, recall = net(image, mask, vertex, vertex_weights) raise TypeError eval_net = DataParallel(EvalWrapper().cuda()) corner_pred = eval_net(seg_pred, vertex_pred).cpu().detach().numpy()[0] camera_matrix = np.array([[572.4114, 0., 325.2611], [0., 573.57043, 242.04899], [0., 0., 1.]]) pose_pred = pnp(points_3d, corner_pred, camera_matrix) projector = Projector() bb8_2d_pred = projector.project(bb8_3d, pose_pred, 'linemod') bb8_2d_gt = projector.project(bb8_3d, pose[0].detach().cpu().numpy(), 'linemod') image = imagenet_to_uint8(image.detach().cpu().numpy())[0] visualize_bounding_box(image[None, ...], bb8_2d_pred[None, None, ...], bb8_2d_gt[None, None, ...]) >>>>>>> 2c722555563b8a77e36b246d82747754cf8dfae7 if __name__ == "__main__": demo()
def demo(): net = Resnet18_8s(ver_dim=vote_num * 2, seg_dim=2) net = NetWrapper(net).cuda() net = DataParallel(net) optimizer = optim.Adam(net.parameters(), lr=train_cfg['lr']) model_dir = os.path.join(cfg.MODEL_DIR, 'cat_demo') load_model(net.module.net, optimizer, model_dir, args.load_epoch) data, points_3d, bb8_3d = read_data() image, mask, vertex, vertex_weights, pose, corner_target = [ d.unsqueeze(0).cuda() for d in data ] # Run the net seg_pred, vertex_pred, loss_seg, loss_vertex, precision, recall = net( image, mask, vertex, vertex_weights) print('vertex_pred.shape') print(vertex_pred.shape) print(' ') print('vertex_pred[0]') print(vertex_pred) print(' ') # Various visualizations #visualize_vertex_field(vertex_pred,vertex_weights, keypointIdx=3) print('asdasdsadas') print(seg_pred.shape, mask.shape) visualize_mask(np.squeeze(seg_pred.cpu().detach().numpy()), mask.cpu().detach().numpy()) rgb = Image.open('data/demo/cat.jpg') img = np.array(rgb) #visualize_overlap_mask(img, np.squeeze(seg_pred.cpu().detach().numpy()), None) # Run the ransac voting eval_net = DataParallel(EvalWrapper2().cuda()) #corner_pred = eval_net(seg_pred, vertex_pred).cpu().detach().numpy()[0] corner_pred, covar = [ x.cpu().detach().numpy()[0] for x in eval_net(seg_pred, vertex_pred) ] print('Keypoint predictions:') print(corner_pred) print(' ') print('covar: ', covar) print(' ') camera_matrix = np.array([[572.4114, 0., 325.2611], [0., 573.57043, 242.04899], [0., 0., 1.]]) # Fit pose to points #pose_pred = pnp(points_3d, corner_pred, camera_matrix) #evaluator = Evaluator() #pose_pred = evaluator.evaluate_uncertainty(corner_pred,covar,pose,'cat',intri_matrix=camera_matrix) def getWeights(covar): cov_invs = [] for vi in range(covar.shape[0]): # For every keypoint if covar[vi, 0, 0] < 1e-6 or np.sum(np.isnan(covar)[vi]) > 0: cov_invs.append(np.zeros([2, 2]).astype(np.float32)) continue cov_inv = np.linalg.inv(scipy.linalg.sqrtm(covar[vi])) cov_invs.append(cov_inv) cov_invs = np.asarray(cov_invs) # pn,2,2 weights = cov_invs.reshape([-1, 4]) weights = weights[:, (0, 1, 3)] return weights weights = getWeights(covar) pose_pred = uncertainty_pnp(corner_pred, weights, points_3d, camera_matrix) print('Predicted pose: \n', pose_pred) print('Ground truth pose: \n', pose[0].detach().cpu().numpy()) print(' ') projector = Projector() bb8_2d_pred = projector.project(bb8_3d, pose_pred, 'linemod') bb8_2d_gt = projector.project(bb8_3d, pose[0].detach().cpu().numpy(), 'linemod') image = imagenet_to_uint8(image.detach().cpu().numpy())[0] visualize_points(image[None, ...], corner_target.detach().cpu().numpy(), pts_pred=corner_pred[None, :, :]) visualize_bounding_box(image[None, ...], bb8_2d_pred[None, None, ...], bb8_2d_gt[None, None, ...])
def demo(idx): net = Resnet18_8s(ver_dim=vote_num * 2, seg_dim=2) net = NetWrapper(net).cuda() net = DataParallel(net) optimizer = optim.Adam(net.parameters(), lr=train_cfg['lr']) model_dir = os.path.join(cfg.MODEL_DIR, "intake_demo") load_model(net.module.net, optimizer, model_dir, -1) data, points_3d, bb8_3d = read_data(idx) #print("BB8_3D: ",bb8_3d) image, mask, vertex, vertex_weights, pose, corner_target = [ d.unsqueeze(0).cuda() for d in data ] seg_pred, vertex_pred, loss_seg, loss_vertex, precision, recall = net( image, mask, vertex, vertex_weights) seg_mask = torch.argmax(seg_pred, 1) visualize_mask(seg_mask) visualize_mask(mask) #visualize_vertex(vertex, vertex_weights) #visualize_hypothesis(image, seg_pred, vertex_pred, corner_target) #visualize_voting_ellipse(image, seg_pred, vertex_pred, corner_target) ############# eval_net = DataParallel(EvalWrapper().cuda()) uncertain_eval_net = DataParallel(UncertaintyEvalWrapper().cuda()) corner_pred = eval_net(seg_pred, vertex_pred).cpu().detach().numpy()[0] net.eval() loss_seg, loss_vertex, precision, recall = [ torch.mean(val) for val in (loss_seg, loss_vertex, precision, recall) ] print("LOSS SEG :", loss_seg, "\nLOSS VERTEX : ", loss_vertex, "\nPRECISION :", precision, '\nRECALL :', recall) ############### #print("Corner Predictions: ",corner_pred) camera_matrix = np.array([[700, 0., 320.], [0., 700, 240.], [0., 0., 1.]]) pose_pred = pnp(points_3d, corner_pred, camera_matrix) projector = Projector() print("Pose prediction :\n", pose_pred) pose_gt = pose[0].detach().cpu().numpy() print("GT Pose :\n", pose[0].detach().cpu().numpy()) s = 0 import math as m for i in range(3): if pose_pred[2][3] < 0: print('NB!') s += (pose_pred[i][3] - pose_gt[i][3])**2 s = m.sqrt(s) print("--->", loss_seg.detach().cpu().numpy(), loss_vertex.detach().cpu().numpy(), precision.detach().cpu().numpy(), recall.detach().cpu().numpy(), s) bb8_2d_pred = projector.project(bb8_3d, pose_pred, 'blender') bb8_2d_gt = projector.project(bb8_3d, pose[0].detach().cpu().numpy(), 'blender') #print(bb8_2d_gt) image = imagenet_to_uint8(image.detach().cpu().numpy())[0] visualize_bounding_box(image[None, ...], bb8_2d_pred[None, None, ...], bb8_2d_gt[None, None, ...])