Function triplet_loss_hard_negative_mining_plus

script/feature/misc.py:399–435

triplet loss with hard negative mining, four cases. inspired by http://www.bmva.org/bmvc/2016/papers/paper119/paper119.pdf section3.3 :param criterion: loss function :param f1: [lvl, B, C, H, W] :param f2: [lvl, B, C, H, W] :return: loss

(f1, f2, margin=1.)

    return loss

def triplet_loss_hard_negative_mining_plus(f1, f2, margin=1.):
    ''' 
    triplet loss with hard negative mining, four cases. inspired by http://www.bmva.org/bmvc/2016/papers/paper119/paper119.pdf section3.3 
    :param criterion: loss function
    :param f1: [lvl, B, C, H, W]
    :param f2: [lvl, B, C, H, W]
    :return:
        loss
    '''
    criterion = nn.TripletMarginLoss(margin=margin, reduction='mean')
    anchor = f1
    anchor_negative = torch.roll(f1, shifts=1, dims=1)
    positive = f2
    negative = torch.roll(f2, shifts=1, dims=1)

    # select in-triplet hard negative, reference: section3.3 
    mse = nn.MSELoss(reduction='mean')
    with torch.no_grad():
        case1 = mse(anchor, negative)
        case2 = mse(positive, anchor_negative)
        case3 = mse(anchor, anchor_negative)
        case4 = mse(positive, negative)
        distance_list = torch.stack([case1,case2,case3,case4])
        loss_case = torch.argmin(distance_list)
    
    # perform anchor swap if necessary
    if loss_case == 0:
        loss = criterion(anchor, positive, negative)
    elif loss_case == 1:
        loss = criterion(positive, anchor, anchor_negative)
    elif loss_case == 2:
        loss = criterion(anchor, positive, anchor_negative)
    elif loss_case == 3:
        loss = criterion(positive, anchor, negative)
    else:
        raise NotImplementedError
    return loss

def perturb_rotation(c2w, theta, phi, psi=0):
    last_row = np.array([[0, 0, 0, 1]])# np.tile(np.array([0, 0, 0, 1]), (1, 1))  # (N_images, 1, 4)