Self-Distilled Depth Refinement with Noisy Poisson Fusion
Authors: Jiaqi Li, Yiran Wang, Jinghong Zheng, Zihao Huang, Ke Xian, Zhiguo Cao, Jianming Zhang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on five benchmarks. SDDR achieves state-of-the-art performance on the commonly-used Middlebury2021 [34], Multiscopic [52], and Hypersim [32]. Meanwhile, since SDDR can establish self-distillation with accurate depth edge representation and guidance on natural scenes, the evaluations on in-the-wild DIML [15] and DIODE [40] datasets showcase our superior generalizability. Analytical experiments demonstrate that these noticeable improvements essentially arise from the strong robustness to the noises. |
| Researcher Affiliation | Collaboration | Jiaqi Li1, Yiran Wang1, Jinghong Zheng1 Zihao Huang1 Ke Xian2 Zhiguo Cao1, Jianming Zhang3 1School of AIA, Huazhong University of Science and Technology 2School of EIC, Huazhong University of Science and Technology 3Adobe Research |
| Pseudocode | No | The paper describes methods using prose and diagrams but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | https://github.com/lijia7/SDDR |
| Open Datasets | Yes | Our training data is sampled from diverse datasets, which can be categorized into synthetic and natural-scene datasets. The synthetic datasets consist of Tartan Air [45], Irs [44], Unreal Stereo4K [39] and MVS-Synth [11]. ... To enhance the generalization to natural scenes, we also sample from four high-resolution real-world datasets, Holopix50K [10], i Bims-1 [16], WSVD [43], and VDW [47]. |
| Dataset Splits | No | The paper mentions 'evaluation datasets' and 'training data' but does not explicitly provide details about a specific validation dataset split used during the training process for hyperparameter tuning. It refers to 'validation set' in the context of generalization testing on the DIODE dataset, which functions as a test set. |
| Hardware Specification | Yes | All training and inference are conducted on a single NVIDIA A6000 GPU. |
| Software Dependencies | No | The paper describes the methodology and implementation details but does not explicitly specify software dependencies with version numbers within the provided text (e.g., specific library versions like PyTorch 1.9, CUDA 11.1). |
| Experiment Setup | Yes | For each epoch, we randomly choose 20,000 images from natural-scene data [10, 47, 43, 16] and 20,000 images from synthetic datasets [45, 44, 39, 11]. For each sample, we adopt similar data processing and augmentation as GBDF [3]. To enhance training stability, we first train Nr for one epoch only with Lgt. In the next two epochs, we involve Lgrad and Lfusion for self-distillation. The a and Nw in Lfusion are set to 0.02 and 4. The learning rate is 1e-4. |