Search papers, labs, and topics across Lattice.
N−1∑oN−1maxs∈[1,N−1](Is,oAcc−IN,oAcc),FM=\frac{1}{N-1}\sum_{o}^{N-1}max_{s\in[1,N-1]}(I_{s,o}^{Acc}-I_{N,o}^{Acc}), (1) where NN denotes the total incremental steps. Is,oAcc/IN,oAccI_{s,o}^{Acc}/I_{N,o}^{Acc} represents the I-AUROC, P-AUROC, or AUPRO score for the oo-th object at the ss-th/N/N-th step. maxs(⋅)max_{s}(\cdot) denotes the maximum score drop for each object across all steps. Implementation Details. In our empirical study protocol, we train a unified model on 6 basic objects from MVTec, Northeastern University
1
0
2
Naively combining unimodal architectures for multimodal anomaly detection makes catastrophic forgetting *worse* – unless you disentangle features with Mamba and filter redundancies with an information bottleneck.
