Computer Vision , AI

● Summary: Continuous convolution filter can be uniformly approximated by a linear combination of high-order differential operators ● Approach highlight Periodic activation function: In implicit neural representation, using the periodic activation function can be useful for models that need to produce the same output for different inputs. Ex) input x, y and output R, G, B values Generate new INR..

● Summary: Template-based new object and occlusion robust 3D pose estimation by contrastive learning ● Approach highlight Local representation(Unseen object performance): masking global feature with binary template mask M to solve the problem of unseen object's cluttered backgrounds Occlusion mask(Occlusion robustness): use a similarity-based occlusion mask O instead of a pooling layer to preser..

● Summary: Active Task Randomization can create tasks for training the skill policies to handle diverse scenarios in sequential manipulation tasks ● Approach highlight Active task Randomization: task sampler propose suitable tasks in the simulation Task Encoder: adaptively estimate the feasibility of sampled tasks Maps each task parameter into a compact embedding($\rightarrow$○) in the replay bu..

● Summary: ChatGPT+Prompt Managing system to use Visual Foundation Models(VFMs) ●Approach highlight Prompt managing of system principles M(P): Transform system principles into a prompt format that ChatGPT can understand. Prompt managing of Foundation Model M(F): help Visual ChatGPT accurately understand and handle the task ● Main Results ● Discussion Is Prompt Engineering a good answer for handl..

● Summary: Multi-view geometry improves object retrieval performance ●Approach highlight Reranking transformer for object retrieval Epipolar Loss and Max-Epipolar Loss: Using epiploar line to utilize multi-view image $𝐴^{12}, 𝐴^{21}$: cross attention map from last transformer 𝕝(𝑖,𝑗): indicator function. if (𝑖,𝑗)on epipolar line then 1, else 0 ● Main Results ● Discussion No significant performanc..

● Summary: stabilizing the loss landscape is crucial to preserve plasticity ● Approach highlight They show abrupt task change can drive instability in optimizers and drive plasticity loss When loss change suddenly, $\hat{m}_{t}$ is updated more aggressively than $\hat{v}_{t}$ and that makes $\hat{u}_{t}$ instability. This simple solution is increasing 𝜖 To understand the impact of optimization m..

● Summary: Get fast and accurate expert models from edge devices with model selection and model retraining strategies ● Approach highlight Model selection: using a gate network to measure the performance of each model zoo. based on this score, select the top-k of model zoo Model retraining: retrain each incoming job from an edge device by 1 epoch and retrain the model with the highest improvemen..

● Summary: Divide and conquer: Class Incremental Learning problem can be divided into within task prediction (WP) and out-of-distribution (OOD) detection and solved by optimizing each part. ● Approach highlight They solve CL by dividing it into TP and WP problems under the following two assumptions: 1. The domains of classes of the same task are disjoint, and 2. The domains of tasks are disjoint..

오늘 리뷰할 논문은 Templates for 3D Object Pose Estimation Revisited: Generalization to New objects and Robustness to Occlusions라는 CVPR 2022년도에 나온 논문입니다. 먼저 들어가기에 앞서 본 논문의 핵심을 말하자면 다음과 같습니다. 1. Template-based object pose estimation model들이 그동안 왜 실패해 왔는지 실험을 통해서 증명하고 positive pair와 negative pair사이의 contrastive learning을 통해 Template-based pose estimation model을 제안합니다. 2. 더 나아가 train data에 없는 new object가 들..

● Summary: Adopts energy-based method models classifier to continual learning to solve class incremental problem ●Approach highlight ○ Energy-based model for classifier: take class y and data x and as input and output is their energy value ○Energy-based models loss function makes minimize energy of positive pair and maximize energy of negative pair ● Main Results ● Discussion ○ Computational cos..