Language-Enhanced Latent Representations for Out-of-Distribution Detection in Autonomous Driving

Presented by Zhenjiang Mao at the Robot Trust for Symbiotic Societies (RTSS) Workshop, ICRA 2024. Also by Dong-You ("Sam") Jhong at UF AI Days 2024 and the NELMS IoT conference 2024.

Out-of-distribution (OOD) detection is essential in autonomous driving, to determine when learning-based components encounter unexpected inputs. Traditional detectors typically use encoder models with fixed settings, thus lacking effective human interaction capabilities. With the rise of large foundation models, multimodal inputs offer the possibility of taking human language as a latent representation, thus enabling language-defined OOD detection. In this paper, we use the cosine similarity of image and text representations encoded by the multimodal model CLIP as a new representation to improve the transparency and controllability of latent encodings used for visual anomaly detection. We compare our approach with existing pre-trained encoders that can only produce latent representations that are meaningless from the user's standpoint. Our experiments on realistic driving data show that the language-based latent representation performs better than the traditional representation of the vision encoder and helps improve the detection performance when combined with standard representations.

1. Language-Enhanced Latent Representations
for Out-of-Distribution Detection
in Autonomous Driving
Zhenjiang Mao, Dong-You Jhong, Ao Wang, Ivan Ruchkin
Department of Electrical and Computer Engineering

Robot Trust for Symbiotic Societies (RTSS) Workshop
05/13/24

1

2. Out-of-Distribution (OOD) Detection
2
Neural Networks (NN)
NN controller
NN safety monitor
Training data
In-distribution data
Classification:
Red Light
Unseen data
Out-of-distribution(OOD) data
Classification:
Green Light

3. Language-enhanced OOD Detection
3
Image
Encoder
distance λ
(w,w’)

…

CLIP (Contrastive Language-Image Pre-Training)
Image
Encoder
v
1
v
2

…

…

…

…

…

v
32

…
Image
Encoder
v
1
v
2

…

…

…

…

…

v
32

v
1
v
2

…

…

…

…

…

v
32

v
1
v
2

…

…

…

…

…

v
32

In distribution
or not?
a rainy dark
night
a sunny clear
day
It drives in
{condition}
Text
Encoder
…
T
1
T
2
v
1
v
2

…

…

…

…

…

v
32

cosine
similarity
I
1
…
Language-based
Encoder
Lang
Encoder
v
1
v
2

…

…

…

…

…

v
32


0.133

0.763
…

0.133

0.763
…

0.133

0.763
…

4. Results
4
OOD types
F1 scores (%)
baseline1 [1]baseline 2 ours ours