Learning End-to-End Scene Flow by Distilling Single Tasks Knowledge
Authors: Filippo Aleotti, Matteo Poggi, Fabio Tosi, Stefano Mattoccia10435-10442
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Exhaustive experiments show that i) DWARF runs at about 10 FPS on a single high-end GPU and about 1 FPS on NVIDIA Jetson TX2 embedded at KITTI resolution, with moderate drop in accuracy compared to 10 deeper models, ii) learning from many distilled samples is more effective than from the few, annotated ones available. 4 Experimental Results We report extensive experiments aimed at assessing the accuracy and performance of DWARF. |
| Researcher Affiliation | Academia | Filippo Aleotti, Matteo Poggi, Fabio Tosi, Stefano Mattoccia Department of Computer Science and Engineering (DISI) University of Bologna, Italy |
| Pseudocode | No | The paper describes the DWARF architecture with block diagrams (Figure 2) and mathematical formulations (e.g., Equation 1 and 2), but it does not include structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any explicit statement about releasing source code or a link to a code repository for the methodology described. |
| Open Datasets | Yes | Flying Things 3D. We set α6 = 0.32, α5 = 0.08, α4 = 0.02, α3 = 0.01 and α2 = 0.005, γ = 0.0004 and crosstask weights to ϵ1 = 1, ϵ2 = 1 and ϵ3 = 0.5. Ground truth values are down-scaled to match the resolution of the level and scaled by a factor of 20, as done by Dosovitskiy et al.; Sun et al. (2015; 2017). KITTI 2015. We fine-tuned the network using the 200 training images from the KITTI Scene Flow (Menze, Heipke, and Geiger 2015) dataset with a batch size of 4 for 50K steps. |
| Dataset Splits | Yes | Flying Things3D. This dataset provides 4248 frames for validation. In Table 1 we report average End-Point-Error (EPE) for the disparity, flow and change (respectively D1, F1 and D2) on 3822 images, obtained by filtering the validation set according to the guidelines. KITTI 2015. For this experiment, we split the KITTI 2015 training set into 160 images for fine-tuning and reserve the last portion of 40 images for validation purposes only. |
| Hardware Specification | Yes | Exhaustive experiments show that i) DWARF runs at about 10 FPS on a single high-end GPU and about 1 FPS on NVIDIA Jetson TX2 embedded at KITTI resolution. Runtime analysis for different variants of DWARF on NVIDIA Jetson TX2 (using Max-Q, Max-P, Max-N configurations) and NVIDIA GTX 1080Ti. We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan Xp GPU used for this research. |
| Software Dependencies | No | The paper mentions the use of Adam optimizer and Leaky ReLU activations, but does not specify software dependencies like deep learning frameworks (e.g., PyTorch, TensorFlow) with their version numbers. |
| Experiment Setup | Yes | Flying Things 3D. We set α6 = 0.32, α5 = 0.08, α4 = 0.02, α3 = 0.01 and α2 = 0.005, γ = 0.0004 and crosstask weights to ϵ1 = 1, ϵ2 = 1 and ϵ3 = 0.5. ... The network has been trained for 1.2M steps with a batch size of 4 randomly selecting crops with size 768 × 384, using Adam optimiser (Kingma and Ba 2014), with β1 = 0.9, β2 = 0.999 and initial learning rate of 10-4, which has been halved after 400K, 600K, 800K and 1M steps. |