personal notes and key points on papers related to DL and AD papers in under 400 words
06/ 2022
- Imbalance Problems in Object Detection: A Review - short summary
- TLDR: Imbalance in object detection comes in various different flavours.
08/ 2021
- Generalizable Pedestrian Detection: The Elephant In The Room - short summary
- TLDR: Pedestrian Detection doesn't generalize well across datasets, even when trained on large corpora of data. Domain shifts are problematic.
07/ 2021
- Deep Depth from Aberration Map - short summary
- TLDR: Physical clues can help monocular depth estimation.
07/ 2021
- Involution: Inverting the Inherence of Convolution for Visual Recognition - short summary
- TLDR: Soon.
06/ 2021
-
Pixel-wise Anomaly Detection in Complex Driving Scenes - short summary
- TLDR: Soon.
-
CSI: Novelty Detection via Contrastive Learning on Distributionally Shifted Instances - short summary
- TLDR: Soon.
05/ 2021
- Involution: Inverting the Inherence of Convolution for Visual Recognition - short summary
- TLDR: Soon.
04/ 2021
- Squeeze-and-Excitation Networks - short summary
- TLDR: Learning global weights for activation maps instead of simply summing them up leads to better performance and fewer parameters.
03/ 2021
-
Deep Unsupervised Anomaly Detection - [short summary](COMING SOON)
- TLDR: Coming soon.
-
Driver Anomaly Detection: A Dataset and Contrastive Learning Approach - [short summary](COMING SOON)
- TLDR: Coming soon.
02/ 2021
- High-Performance Large-Scale Image Recognition Without Normalization - [short summary](COMING SOON)
- TLDR: Coming soon.
01/ 2021
-
Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference - [short summary](COMING SOON)
- TLDR: The most iconic paper for neural network acceleration to date.
-
Mixed Precision Training - [short summary](COMING SOON)
- TLDR: Three techniques for neural network compression for faster inference and training speed: master copying FP32 weights, loss scaling and FP16 calculation/FP32 accumulation.
-
What Do Compressed Deep Neural Networks Forget? - short summary
- TLDR: Compression disproportionately impacts model performance on the underrepresented long-tail of the data distribution.
12/ 2020
-
Detection and Retrieval of Out-of-Distribution Objects in Semantic Segmentation - short summary
- TLDR: Yet another interesting method for uncertainty quantification in semantic segmentation NNs.
-
Attribution Preservation in Network Compression for Reliable Network Interpretation - short summary
- TLDR: Compressing neural networks for inference is in conflict with information attribution, which can be used for explanation of network outputs in safety-critical domains.
-
STDP-based spiking deep convolutional neural networks for object recognition - short summary
- TLDR: Unlike "classical" neural networks, bio-inspired spiking neural net make heavy use of timing for both learning and inference.
11/ 2020
- Towards spike-based machine intelligence with neuromorphic computing - short summary
- TLDR: The brain has spiking neurons, current AI does not. Neuromorphic computing utilizes algorithms and software–hardware codesign to realize spiking neural networks.
09 / 2020
- The Fishyscapes Benchmark:
Measuring Blind Spots in Semantic Segmentation - short summary
- TLDR: A dataset and benchmark for the critical task of handling unknown objects in semantic segmentations.
08 / 2020
- Deep Compression: Compressing Deep Neural Networks with pruning, trained quantization and Huffman coding - short summary
- TLDR: One of the very first papers on modern neural network compression. Iconic.
07 / 2020
-
Training End-to-End Analog Neural Networks with Equilibrium Propagation - short summary
- TLDR: A proof that end-to-end training of analog neural networks works in practice.
-
Disentangling Monocular 3D Object Detection - short summary
- TLDR: Interesting. Deriving 3D bounding boxes from 2D images, powered by an improved loss function.
05 / 2020
- EfficientDet: Scalable and Efficient Object Detection - short summary
- TLDR: A clever architecture for multi-scale feature aggregation vastly improves the state of the art in object detection and semantic segmentation. Awesome.
03 / 2020
- Class-Balanced Loss Based on Effective Number of Samples - short summary
- TLDR: Vanilla reweighting or resampling strategies suck at representing long-tailed distributions. Geometrical representations of data can help with balancing labels for deep learning.
02 / 2020
- MeteorNet: Deep Learning on Dynamic 3D Point Cloud Sequences - short summary
- TLDR: Incorporating spatiotemporal relationships by design improves performance. Based on PointNet.
- Performance Evaluation of Object Tracking Algorithms - short summary
- TLDR: Performance for video tracking algorithms can be evaluated with various metrices with different strengths and weaknesses.
01 / 2020
- FuseSeg: LiDAR Point Cloud Segmentation Fusing Multi-Modal Data - short summary
- TLDR: Cool high-level sensor fusion method of RGB and LiDAR data using learned point correspondences.
- StarNet: Targeted Computation for
Object Detection in Point Clouds - short summary
- TLDR: Yet another method for non-volumetric 3D object detection. State-of-the-art on Waymo Open.
12 / 2019
- PointRend: Image Segmentation as Rendering - short summary
- TLDR: In semantic segmentation, focusing on areas of high uncertainty along with adaptively upsampling seems to be a winning combination.
- PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation - short summary
- TLDR: Performing 3D semantic segmentation on unordered point sets without 3D convolutions.
- Adversarial Examples Improve Image Recognition - [short summary](coming soon)
- TLDR: Image classification is already at human level, but it can further be improved with adversarial examples.
11 / 2019
-
Rethinking Atrous Convolution for Semantic Image Segmentation (DeepLab) - short summary
- TLDR: Atrous convolutions resolves the issue of intentional information loss through maxpooling. Also, control the field-of-view of filters more efficiently.
-
Gaussian YOLOv3: An Accurate and Fast Object Detector Using Localization Uncertainty for Autonomous Driving - short summary
- TLDR: Modelling the uncertainty in bounding box predictions of YOLOv3 as a Gaussian parameter.
10 / 2019
-
The ML Test Score: A Rubric for ML Production Readiness and Technical Debt Reduction - short summary
- TLDR: Bringing software engineering best practices to deep learning.
-
VoxelNet: End-to-End Learning for Point Cloud Based 3D Object Detection - short summary
- TLDR: End-to-end deep learning on point clouds requires special operations, in this case Stacked Voxel Feature Encoding.
09 / 2019
- A critique of pure learning: what artificial neural nets can learn from animal brains - short summary
08 / 2019
- Events-to-Video: Bringing Modern Computer Vision to Event Cameras - [short summary](coming soon)
- Fast Online Object Tracking and Segmentation: A Unifying Approach - short summary
- Hybrid Task Cascade for Instance Segmentation - short summary
07 / 2019
- LaserNet: An Efficient Probabilistic 3D Object Detector for Autonomous Driving - coming soon
- A Survey of Deep Learning-based Object Detection - short summary
- TLDR: Great resource on 2D and 3D object detection and corresponding datasets.
06 / 2019
-
VoxelNet: End-to-End Learning for Point Cloud Based 3D Object Detection - short summary
-
Visualizing the Loss Landscape of Neural Nets - short summary
- TLDR: Neat method to learn about the shape of the loss surface in deep neural networks.
-
Adversarial Examples Are Not Bugs, They Are Features - short summary
- TLDR: Dividing features of an image classifier into robust and non-robust helps causally determine effects of adversarial examples
05 / 2019
- GAN Dissection: Visualizing and Understanding Generative Adversarial Networks - short summary
- Semantic Segmentation using Adversarial Networks - short summary
04 / 2019
11 / 2018
10 / 2018
- BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding - short summary
- TLDR: Transformers are the designated successor of LSTMs and outperform them on many tasks.
12 / 2017
-
Dynamic Routing Between Capsules - short summary
- TLDR Introducing a prior with regards to neuron grouping makes various learning tasks easier, but computationally challenging. 11 / 2017
-
Video Object Segmentation with Re-identification PDF - short summary
10 / 2017
- Learning Video Object Segmentation from Static Images - short summary
- TLDR Introducing a prior with regards to temporal correlation lets you do semi-supervised-style labelling.
09 / 2017
- Training RNNs as Fast as CNNs - short summary TLDR: Temporal dependencies slow down RNN training. This can be remedied using highway connections.
- Video Frame Interpolation via Adaptive Separable Convolution ARXIV - short summary
- Speech recognition with deep recurrent neural networks ARXIV - short summary
- Brain Tumor Segmentation with Deep Neural Networks ARXIV - short summary
03/ 2021
- Driver Anomaly Detection: A Dataset and Contrastive Learning Approach - [short summary](COMING SOON)
- TLDR: Coming soon.
12 / 2020
- Tire and Vehicle Dynamics - [short summary](Coming soon.)
- TLDR: Coming soon.
11/ 2020
- Attribution Preservation in Network Compression for Reliable Network Interpretation - short summary
- TLDR: Compressing neural networks for inference is in conflict with information attribution, which can be used for explanation of network outputs in safety-critical domains.
01 / 2020
- Night-Time Vehicle Sensing in Far Infrared Image with Deep Learning - short summary
- TLDR: Far infrared automotive sensors offer night vision and can be used for learning-based object detection.
- Next Generation Radar Sensors in Automotive
Sensor Fusion Systems - short summary
- TLDR: Radars are cool and provide unique sensoric capabilities indespensable for successful fully AD.
12 / 2019
- Road detection based on the fusion of Lidar and image data - short summary
- TLDR: Fuse image and 3D LiDaR data by first projecting the point clouds onto 2D monocular images, then perform road detection using a conditional random field.
- Certifiability of Deep Learning models in security critical industries such as automotive short summary
- TLDR: Large-scale deployment of deep learning models needs certification from authorities, which does not exist yet.
11 / 2019
- Stereo R-CNN based 3D Object Detection for Autonomous Driving - short summary
- TLDR: End-to-end learning of an 3D object detection model using stereo camera inputs.
10 / 2019
- Autonomous Vehicle Implementation Prediction - short summary
- TLDR: Autonomous cars might cause enormous benefits to society, but we're not there yet. From an city planning perspective.
08 / 2019
- DynaNet: Neural Kalman Dynamical Model for Motion Estimation and Prediction - short summary
- A Multimodal Vision Sensor for Autonomous Driving - short summary
- Deep Sensor Fusion for Real-Time Odometry Estimation - short summary
- TLDR: Theoretically, self-localization can be enhanced considerably using 2D LiDAR-scanners. The practicability of the approach is to be questionned.
07 / 2019
- ChauffeurNet: Learning to Drive by Imitating the Best and Synthesizing the Worst - short summary
- TLDR: Imitation learning (what children do) can work for autonomous driving, but only if you provide data on malicious situations such as going off the road / collisions (e.g. via simulation).
06 / 2019
-
Self-Driving Cars: A Survey - short summary
- TLDR: Self-driving car architectures share many similar features. This paper gathers literature on all the subsystems necessary for it to run.
-
LaserNet: An Efficient Probabilistic 3D Object Detector for Autonomous Driving - [short summary](coming soon)