Result details - ETH3D

Submitted anonymously.

Submission data

Full name	StereoMatchingWithCustomDatasetPretrainingAndBenchmark-SpecificFine-Tuning
Description	This submission evaluates the effect of pretraining data on stereo matching performance. Two training strategies are compared. In the first strategy, the model is trained from scratch on a custom dataset and subsequently fine-tuned on ETH3D, KITTI, and Middlebury. In the second strategy, the model is trained from scratch on Scene Flow and then fine-tuned on the same benchmark datasets. Fine-tuning is performed whenever a dedicated training split is available, even when final evaluation is conducted through the benchmark platform. The objective is to assess whether pretraining on the custom dataset provides better transferability and downstream performance than pretraining on Scene Flow.
Parameters	Experiment 1: - Train from scratch on CustomDataset - Fine-tune on ETH3D - Fine-tune on KITTI - Fine-tune on Middlebury Experiment 2: - Train from scratch on Scene Flow - Fine-tune on ETH3D - Fine-tune on KITTI - Fine-tune on Middlebury Note: Fine-tuning is performed whenever a separate training split is available. Expected outcome: models pretrained on CustomDataset achieve better performance after fine-tuning than models pretrained on Scene Flow.
Programming language(s)	Python, PyTorch, CUDA
Hardware	A100
Submission creation date	11 Jun, 2026
Last edited	11 Jun, 2026

High-res multi-view results

Info	all	high-res multi-view	indoor	outdoor	courty.	delive.	electro	facade	kicker	meadow	office	pipes	playgr.	relief	relief.	terrace	terrai.
No results yet.

Low-res many-view results

Info	all	low-res many-view	indoor	outdoor	lakeside	sand box	storage room	storage room 2	tunnel
No results yet.

Low-res two-view results

Info	all	lakes. 1l	lakes. 1s	sand box 1l	sand box 1s	stora. room 1l	stora. room 1s	stora. room 2l	stora. room 2s	stora. room 2 1l	stora. room 2 1s	stora. room 2 2l	stora. room 2 2s	stora. room 3l	stora. room 3s	tunnel 1l	tunnel 1s	tunnel 2l	tunnel 2s	tunnel 3l	tunnel 3s
							55.43	27.33	29.43	20.86	24.21	22.72	23.17	24.65	28.47	13.68	16.54	12.29	14.47	16.02

SLAM results

		all	boxes	boxes dark	buddha	cables 4	cables 5	desk 1	desk 2	desk changing 2	desk dark 1	desk dark 2	desk global light changes	desk ir light	dino	drone	foreground occlusion	helmet	kidnap 2	lamp	large loop 2	large loop 3	large non loop	motion 2	motion 3	motion 4	planar 1	reflective 2	scale change	table 1	table 2	table 5	table 6	table global light changes	table local light changes	table scene	trashbin
	Method	Info
No results yet.