In order to reveal the rationale behind model predictions, many works have exploited providing explanations in various forms. Recently, to further guarantee readability, more and more works turn to generate sentence-level human language explanations. However, current works pursuing sentence-level explanations rely heavily on annotated training data, which limits the development of interpretability to only a few tasks. As far as we know, this paper is the first to explore this problem smoothly from weak-supervised learning to unsupervised learning. Besides, we also notice the high latency of autoregressive sentence-level explanation generation, which leads to asynchronous interpretability after prediction. Therefore, we propose a non-autoregressive interpretable model to facilitate parallel explanation generation and simultaneous prediction. Through extensive experiments on Natural Language Inference task and Spouse Prediction task, we find that users are able to train classifiers with comparable performance $10-15\times$ faster with parallel explanation generation using only a few or no annotated training data.
In this paper, we are interested in understanding self-supervised pretraining through studying the capability that self-supervised representation pretraining methods learn part-aware representations. The study is mainly motivated by that random views, used in contrastive learning, and random masked (visible) patches, used in masked image modeling, are often about object parts. We explain that contrastive learning is a part-to-whole task: the projection layer hallucinates the whole object representation from the object part representation learned from the encoder, and that masked image modeling is a part-to-part task: the masked patches of the object are hallucinated from the visible patches. The explanation suggests that the self-supervised pretrained encoder is required to understand the object part. We empirically compare the off-the-shelf encoders pretrained with several representative methods on object-level recognition and part-level recognition. The results show that the fully-supervised model outperforms self-supervised models for object-level recognition, and most self-supervised contrastive learning and masked image modeling methods outperform the fully-supervised method for part-level recognition. It is observed that the combination of contrastive learning and masked image modeling further improves the performance.
While dynamic Neural Radiance Fields (NeRF) have shown success in high-fidelity 3D modeling of talking portraits, the slow training and inference speed severely obstruct their potential usage. In this paper, we propose an efficient NeRF-based framework that enables real-time synthesizing of talking portraits and faster convergence by leveraging the recent success of grid-based NeRF. Our key insight is to decompose the inherently high-dimensional talking portrait representation into three low-dimensional feature grids. Specifically, a Decomposed Audio-spatial Encoding Module models the dynamic head with a 3D spatial grid and a 2D audio grid. The torso is handled with another 2D grid in a lightweight Pseudo-3D Deformable Module. Both modules focus on efficiency under the premise of good rendering quality. Extensive experiments demonstrate that our method can generate realistic and audio-lips synchronized talking portrait videos, while also being highly efficient compared to previous methods.
Masked image modeling (MIM) learns visual representation by masking and reconstructing image patches. Applying the reconstruction supervision on the CLIP representation has been proven effective for MIM. However, it is still under-explored how CLIP supervision in MIM influences performance. To investigate strategies for refining the CLIP-targeted MIM, we study two critical elements in MIM, i.e., the supervision position and the mask ratio, and reveal two interesting perspectives, relying on our developed simple pipeline, context autodecoder with CLIP target (CAE v2). Firstly, we observe that the supervision on visible patches achieves remarkable performance, even better than that on masked patches, where the latter is the standard format in the existing MIM methods. Secondly, the optimal mask ratio positively correlates to the model size. That is to say, the smaller the model, the lower the mask ratio needs to be. Driven by these two discoveries, our simple and concise approach CAE v2 achieves superior performance on a series of downstream tasks. For example, a vanilla ViT-Large model achieves 81.7% and 86.7% top-1 accuracy on linear probing and fine-tuning on ImageNet-1K, and 55.9% mIoU on semantic segmentation on ADE20K with the pre-training for 300 epochs. We hope our findings can be helpful guidelines for the pre-training in the MIM area, especially for the small-scale models.
While various knowledge distillation (KD) methods in CNN-based detectors show their effectiveness in improving small students, the baselines and recipes for DETR-based detectors are yet to be built. In this paper, we focus on the transformer decoder of DETR-based detectors and explore KD methods for them. The outputs of the transformer decoder lie in random order, which gives no direct correspondence between the predictions of the teacher and the student, thus posing a challenge for knowledge distillation. To this end, we propose MixMatcher to align the decoder outputs of DETR-based teachers and students, which mixes two teacher-student matching strategies, i.e., Adaptive Matching and Fixed Matching. Specifically, Adaptive Matching applies bipartite matching to adaptively match the outputs of the teacher and the student in each decoder layer, while Fixed Matching fixes the correspondence between the outputs of the teacher and the student with the same object queries, with the teacher's fixed object queries fed to the decoder of the student as an auxiliary group. Based on MixMatcher, we build \textbf{D}ecoder \textbf{D}istillation for \textbf{DE}tection \textbf{TR}ansformer (D$^3$ETR), which distills knowledge in decoder predictions and attention maps from the teachers to students. D$^3$ETR shows superior performance on various DETR-based detectors with different backbones. For example, D$^3$ETR improves Conditional DETR-R50-C5 by $\textbf{7.8}/\textbf{2.4}$ mAP under $12/50$ epochs training settings with Conditional DETR-R101-C5 as the teacher.
We present a strong object detector with encoder-decoder pretraining and finetuning. Our method, called Group DETR v2, is built upon a vision transformer encoder ViT-Huge~\cite{dosovitskiy2020image}, a DETR variant DINO~\cite{zhang2022dino}, and an efficient DETR training method Group DETR~\cite{chen2022group}. The training process consists of self-supervised pretraining and finetuning a ViT-Huge encoder on ImageNet-1K, pretraining the detector on Object365, and finally finetuning it on COCO. Group DETR v2 achieves $\textbf{64.5}$ mAP on COCO test-dev, and establishes a new SoTA on the COCO leaderboard https://paperswithcode.com/sota/object-detection-on-coco
Detection Transformer (DETR) relies on One-to-One assignment, i.e., assigning one ground-truth object to only one positive object query, for end-to-end object detection and lacks the capability of exploiting multiple positive object queries. We present a novel DETR training approach, named {\em Group DETR}, to support Group-wise One-to-Many assignment. We make simple modifications during training: (i) adopt $K$ groups of object queries; (ii) conduct decoder self-attention on each group of object queries with the same parameters; (iii) perform One-to-One label assignment for each group, leading to $K$ positive object queries for each ground-truth object. In inference, we only use one group of object queries, making no modifications to DETR architecture and processes. We validate the effectiveness of the proposed approach on DETR variants, including Conditional DETR, DAB-DETR, DN-DETR, and DINO. Code will be available.
In this paper, we are interested in Detection Transformer (DETR), an end-to-end object detection approach based on a transformer encoder-decoder architecture without hand-crafted postprocessing, such as NMS. Inspired by Conditional DETR, an improved DETR with fast training convergence, that presented box queries (originally called spatial queries) for internal decoder layers, we reformulate the object query into the format of the box query that is a composition of the embeddings of the reference point and the transformation of the box with respect to the reference point. This reformulation indicates the connection between the object query in DETR and the anchor box that is widely studied in Faster R-CNN. Furthermore, we learn the box queries from the image content, further improving the detection quality of Conditional DETR still with fast training convergence. In addition, we adopt the idea of axial self-attention to save the memory cost and accelerate the encoder. The resulting detector, called Conditional DETR V2, achieves better results than Conditional DETR, saves the memory cost and runs more efficiently. For example, for the DC$5$-ResNet-$50$ backbone, our approach achieves $44.8$ AP with $16.4$ FPS on the COCO $val$ set and compared to Conditional DETR, it runs $1.6\times$ faster, saves $74$\% of the overall memory cost, and improves $1.0$ AP score.
Neural Radiance Field (NeRF) has emerged as a compelling method to represent 3D objects and scenes for photo-realistic rendering. However, its implicit representation causes difficulty in manipulating the models like the explicit mesh representation. Several recent advances in NeRF manipulation are usually restricted by a shared renderer network, or suffer from large model size. To circumvent the hurdle, in this paper, we present an explicit neural field representation that enables efficient and convenient manipulation of models. To achieve this goal, we learn a hybrid tensor rank decomposition of the scene without neural networks. Motivated by the low-rank approximation property of the SVD algorithm, we propose a rank-residual learning strategy to encourage the preservation of primary information in lower ranks. The model size can then be dynamically adjusted by rank truncation to control the levels of detail, achieving near-optimal compression without extra optimization. Furthermore, different models can be arbitrarily transformed and composed into one scene by concatenating along the rank dimension. The growth of storage cost can also be mitigated by compressing the unimportant objects in the composed scene. We demonstrate that our method is able to achieve comparable rendering quality to state-of-the-art methods, while enabling extra capability of compression and composition. Code will be made available at \url{https://github.com/ashawkey/CCNeRF}.
Semantic scene reconstruction from point cloud is an essential and challenging task for 3D scene understanding. This task requires not only to recognize each instance in the scene, but also to recover their geometries based on the partial observed point cloud. Existing methods usually attempt to directly predict occupancy values of the complete object based on incomplete point cloud proposals from a detection-based backbone. However, this framework always fails to reconstruct high fidelity mesh due to the obstruction of various detected false positive object proposals and the ambiguity of incomplete point observations for learning occupancy values of complete objects. To circumvent the hurdle, we propose a Disentangled Instance Mesh Reconstruction (DIMR) framework for effective point scene understanding. A segmentation-based backbone is applied to reduce false positive object proposals, which further benefits our exploration on the relationship between recognition and reconstruction. Based on the accurate proposals, we leverage a mesh-aware latent code space to disentangle the processes of shape completion and mesh generation, relieving the ambiguity caused by the incomplete point observations. Furthermore, with access to the CAD model pool at test time, our model can also be used to improve the reconstruction quality by performing mesh retrieval without extra training. We thoroughly evaluate the reconstructed mesh quality with multiple metrics, and demonstrate the superiority of our method on the challenging ScanNet dataset.