Files
logseq/pages/ICSE2023-Poster-6.md
T
2025-06-05 22:07:12 +02:00

2.4 KiB

type:: Referaggi date:: 19-02-2023 - 00:28 external-link:: file:: icse2023-poster-paper6.pdf

- [[Highlights]]
	- ((63f1cbbc-4a1a-491e-8465-8b576e2c949d))
		- What does it mean?
	- ((63f1cfa9-832f-49b9-b6b2-13d64af880a5))
		- This part is unclear. How do the image creation work, in particular how the objects to be added for generating OOD images are decided/identified/created and merged with a same background?
			- THe following is not clear to me:
				- ((63f1d03d-00a5-4dcc-81e9-361e5f340323))
	- What to do with the created images?
	- ((63f1d062-fcb8-41c9-9207-95dfb9626465))
		- Erroneous with respect to what?
		- ((63f1d1ec-bc72-4ae3-b565-e46e612e9dce))
			- that's obscure
- [[Reviews]]
	- The paper presents DISTROFAIR, an approach to automatically generate images to detect class-level fairness violations of existing image recognition software systems. The generation of testing images is based on three mutation operators, i.e., object deletion, object insertion, and object rotation.
	- The paper is about an interesting and relevant problem. I have some doubts concerning the DISTROFAIR approach, as is presented in Fig. 1:
		- As far as I understood, the employed clustering technique plays a crucial role in the overall process. Misplacing objects in the wrong clusters can affect the subsequent phases of the approach. It is important to clarify this aspect and mention how the authors have identified and configured the employed clustering technique.
		- The way the mutation operators work could be clearer to me. The system can add, remove, or rotate objects from the initial image employing dedicated operators. However, authors should clarify what's the definition of semantic preservation relation so that it is possible to automatically decide when the original and the mutation images are in such a relation
		- Following the previous point, it is necessary to clarify the criteria used to decide the parts of the original images that need to be unchanged. For instance, the authors should mention how the "background" of the original image is decided so that mutations are operated on top of it
		- The authors should also clarify how the objects to be added/removed/rotated are decided and used to mutate the original figure.
		- The strengths and the limitations of DISTROFAIR need to be better discussed, e.g., by considering existing baselines, which are currently neglected.