RelTR - Intro to ML Project

Overview

In this project, we summarize the paper RelTR: Relational Transformer for scene graph generation and validate a claim made by the authors.

Overall architecture of RelTR[1]

Claim Validation

The authors claim RelTR[1] has lower number of parameters compared to other scene graph generation models such as FCSGG[2].

Performance and parameter comparison of RelTR with other models[]

FCSGG[] has multiple configurations with varying backbone architectures. Since the authors of RelTR do not specify which configuration is used and RelTR uses the ResNet50 as a backbone, we decided it would only be fair to test when both architectures use the same backbone.

Claim Validation: results and discussion

Model	Parameters (M) ↓
RelTR	67.9
FCSGG	26.4

Table comparing the number of parameters (M) for the models RelTR[1] and FCSGG[2]. ↓ lower is better

We observe that RelTR has more parameters than FCSGG when both architectures use ResNet50 as their backbone.

Result: The claim RelTr has lesser parameters than FCSGG is False.

References

• [1] https://github.com/yrcong/RelTR
• [2] https://github.com/liuhengyue/fcsgg

Authors

• Shubham Gupta
• Karan Sharma