Biplob Debnath NEC Labs America

Biplob Debnath

Senior Researcher

Integrated Systems

Posts

A Perspective on Deep Vision Performance with Standard Image and Video Codecs

/in /by

Resource-constrained hardware such as edge devices or cell phones often rely on cloud servers to provide the required computational resources for inference in deep vision models. However transferring image and video data from an edge or mobile device to a cloud server requires coding to deal with network constraints. The use of standardized codecs such as JPEG or H.264 is prevalent and required to ensure interoperability. This paper aims to examine the implications of employing standardized codecs within deep vision pipelines. We find that using JPEG and H.264 coding significantly deteriorates the accuracy across a broad range of vision tasks and models. For instance strong compression rates reduce semantic segmentation accuracy by more than 80% in mIoU. In contrast to previous findings our analysis extends beyond image and action classification to localization and dense prediction tasks thus providing a more comprehensive perspective.

Read more

Deep Video Codec Control for Vision Models

/in /by

Standardized lossy video coding is at the core of almost all real-world video processing pipelines. Rate control is used to enable standard codecs to adapt to different network bandwidth conditions or storage constraints. However standard video codecs (e.g. H.264) and their rate control modules aim to minimize video distortion w.r.t. human quality assessment. We demonstrate empirically that standard-coded videos vastly deteriorate the performance of deep vision models. To overcome the deterioration of vision performance this paper presents the first end-to-end learnable deep video codec control that considers both bandwidth constraints and downstream deep vision performance while adhering to existing standardization. We demonstrate that our approach better preserves downstream deep vision performance than traditional standard video coding.

Read more

ViTA: An Efficient Video-to-Text Algorithm using VLM for RAG-based Video Analysis System

/in /by

Retrieval-augmented generation (RAG) is used in natural language processing (NLP) to provide query-relevant information in enterprise documents to large language models (LLMs). Such enterprise context enables the LLMs to generate more informed and accurate responses. When enterprise data is primarily videos AI models like vision language models (VLMs) are necessary to convert information in videos into text. While essential this conversion is a bottleneck especially for large corpus of videos. It delays the timely use of enterprise videos to generate useful responses. We propose ViTA a novel method that leverages two unique characteristics of VLMs to expedite the conversion process. As VLMs output more text tokens they incur higher latency. In addition large (heavyweight) VLMs can extract intricate details from images and videos but they incur much higher latency per output token when compared to smaller (lightweight) VLMs that may miss details. To expedite conversion ViTA first employs a lightweight VLM to quickly understand the gist or overview of an image or a video clip and directs a heavyweight VLM (through prompt engineering) to extract additional details by using only a few (preset number of) output tokens. Our experimental results show that ViTA expedites the conversion time by as much as 43% without compromising the accuracy of responses when compared to a baseline system that only uses a heavyweight VLM.

Read more

LeanContext: Cost-efficient Domain-specific Question Answering Using LLMs

/in /by

Question-answering (QA) is a significant application of Large Language Models (LLMs), shaping chatbot capabilities across healthcare, education, and customer service. However, widespread LLM integration presents a challenge for small businesses due to the high expenses of LLM API usage. Costs rise rapidly when domain-specific data (context) is used alongside queries for accurate domain-specific LLM responses. Extracting context from domain-specific data is implemented by a Retrieval Augmented Generation (RAG) approach. One option is to summarize the RAG context by using LLMs and reduce the context. However, this can also filter out useful information that is necessary to answer some domain-specific queries. In this paper, we shift from human-oriented summarizers to AI model-friendly summaries. Our approach, LeanContext, efficiently extracts k key sentences from the context that are closely aligned with the query. The choice of k is neither static nor random; we introduce a reinforcement learning technique that dynamically determines k based on the query and context. The rest of the less important sentences are either reduced using a free open-source text reduction method or eliminated. We evaluate LeanContext against several recent query-aware and query-unaware context reduction approaches on prominent datasets (arxiv papers and BBC news articles, NarrativeQA). Despite cost reductions of 37.29% to 67.81%, LeanContext’s ROUGE-1 score decreases only by 1.41% to 2.65% compared to a baseline that retains the entire context (no summarization). LeanContext stands out for its ability to provide precise responses, outperforming competitors by leveraging open-source summarization techniques. Human evaluations of the responses further confirm and validate this superiority. Additionally, if open-source pre-trained LLM-based summarizers are used to reduce context (into human consumable summaries), LeanContext can further modify the reduced context to enhance the accuracy (ROUGE-1 score) by 13.22% to 24.61%.

Read more

iRAG: An Incremental Retrieval Augmented Generation System for Videos

/in /by

Retrieval augmented generation (RAG) systems combine the strengths of language generation and information retrieval to power many real-world applications like chatbots. Use of RAG for combined understanding of multimodal data such as text, images and videos is appealing but two critical limitations exist: one-time, upfront capture of all content in large multimodal data as text descriptions entails high processing times, and not all information in the rich multimodal data is typically in the text descriptions. Since the user queries are not known apriori, developing a system for multimodal to text conversion and interactive querying of multimodal data is challenging.To address these limitations, we propose iRAG, which augments RAG with a novel incremental workflow to enable interactive querying of large corpus of multimodal data. Unlike traditional RAG, iRAG quickly indexes large repositories of multimodal data, and in the incremental workflow, it uses the index to opportunistically extract more details from select portions of the multimodal data to retrieve context relevant to an interactive user query. Such an incremental workflow avoids long multimodal to text conversion times, overcomes information loss issues by doing on-demand query-specific extraction of details in multimodal data, and ensures high quality of responses to interactive user queries that are often not known apriori. To the best of our knowledge, iRAG is the first system to augment RAG with an incremental workflow to support efficient interactive querying of large, real-world multimodal data. Experimental results on real-world long videos demonstrate 23x to 25x faster video to text ingestion, while ensuring that quality of responses to interactive user queries is comparable to responses from a traditional RAG where all video data is converted to text upfront before any querying.

Read more

Differentiable JPEG: The Devil is in The Details

/in /by

JPEG remains one of the most widespread lossy image coding methods. However, the non-differentiable nature of JPEG restricts the application in deep learning pipelines. Several differentiable approximations of JPEG have recently been proposed to address this issue. This paper conducts a comprehensive review of existing diff. JPEG approaches and identifies critical details that have been missed by previous methods. To this end, we propose a novel diff. JPEG approach, overcoming previous limitations. Our approach is differentiable w.r.t. the input image, the JPEG quality, the quantization tables, and the color conversion parameters. We evaluate the forward and backward performance of our diff. JPEG approach against existing methods. Additionally, extensive ablations are performed to evaluate crucial design choices. Our proposed diff. JPEG resembles the (non-diff.) reference implementation best, significantly surpassing the recent-best diff. approach by 3.47dB (PSNR) on average. For strong compression rates, we can even improve PSNR by 9.51dB. Strong adversarial attack results are yielded by our diff. JPEG, demonstrating the effective gradient approximation. Our code is available at https://github.com/necla-ml/Diff-JPEG.

Read more

Deep Video Codec Control

/in /by

Deep Video Codec Control Lossy video compression is commonly used when transmitting and storing video data. Unified video codecs (e.g., H.264 or H.265) remain the emph(Unknown sysvar: (de facto)) standard, despite the availability of advanced (neural) compression approaches. Transmitting videos in the face of dynamic network bandwidth conditions requires video codecs to adapt to vastly different compression strengths. Rate control modules augment the codec’s compression such that bandwidth constraints are satisfied and video distortion is minimized. While, both standard video codes and their rate control modules are developed to minimize video distortion w.r.t. human quality assessment, preserving the downstream performance of deep vision models is not considered. In this paper, we present the first end-to-end learnable deep video codec control considering both bandwidth constraints and downstream vision performance, while not breaking existing standardization. We demonstrate for two common vision tasks (semantic segmentation and optical flow estimation) and on two different datasets that our deep codec control better preserves downstream performance than using 2-pass average bit rate control while meeting dynamic bandwidth constraints and adhering to standardizations.

Read more

FactionFormer: Context-Driven Collaborative Vision Transformer Models for Edge Intelligence

/in /by

Edge Intelligence has received attention in the recent times for its potential towards improving responsiveness, reducing the cost of data transmission, enhancing security and privacy, and enabling autonomous decisions by edge devices. However, edge devices lack the power and compute resources necessary to execute most Al models. In this paper, we present FactionFormer, a novel method to deploy resource-intensive deep-learning models, such as vision transformers (ViT), on resource-constrained edge devices. Our method is based on a key observation: edge devices are often deployed in settings where they encounter only a subset of the classes that the resource intensive Al model is trained to classify, and this subset changes across deployments. Therefore, we automatically identify this subset as a faction, devise on-the fly a bespoke resource-efficient ViT called a modelette for the faction and set up an efficient processing pipeline consisting of a modelette on the device, a wireless network such as 5G, and the resource-intensive ViT model on an edge server, all of which work collaboratively to do the inference. For several ViT models pre-trained on benchmark datasets, FactionFormer’s modelettes are up to 4× smaller than the corresponding baseline models in terms of the number of parameters, and they can infer up to 2.5× faster than the baseline setup where every input is processed by the resource-intensive ViT on the edge server. Our work is the first of its kind to propose a device-edge collaborative inference framework where bespoke deep learning models for the device are automatically devised on-the-fly for most frequently encountered subset of classes.

Read more

Efficient Compression Method for Roadside LiDAR Data

/in /by

Roadside LiDAR (Light Detection and Ranging) sensors are recently being explored for intelligent transportation systems aiming at safer and faster traffic management and vehicular operations. A key challenge in such systems is to efficiently transfer massive point-cloud data from the roadside LiDAR devices to the edge connected through a 5G network for real-time processing. In this paper, we consider the problem of compressing roadside (i.e. static) LiDAR data in real-time that provides a unique condition unexplored by current methods. Existing point-cloud compression methods assume moving LiDARs (that are mounted on vehicles) and do not exploit spatial consistency across frames over time.To this end, we develop a novel grouped wavelet technique for static roadside LiDAR data compression (i.e. SLiC). Our method compresses LiDAR data both spatially and temporally using a kd-tree data structure based on Haar wavelet coefficients. Experimental results show that SLiC can compress up to 1.9× more effectively than the state-of-the-art compression method can do. Moreover, SLiC is computationally more efficient to achieve 2× improvement in bandwidth usage over the best alternative. Even with this impressive gain in communication and storage efficiency, SLiC retains down-the-pipeline application’s accuracy.

Read more

Cosine Similarity based Few-Shot Video Classifier with Attention-based Aggregation

/in /by

Meta learning algorithms for few-shot video recognition use complex, episodic training but they often fail to learn effective feature representations. In contrast, we propose a new and simpler few-shot video recognition method that does not use meta-learning, but its performance compares well with the best meta-learning proposals. Our new few-shot video classification pipeline consists of two distinct phases. In the pre-training phase, we learn a good video feature extraction network that generates a feature vector for each video. After a sparse sampling strategy selects frames from the video, we generate a video feature vector from the sampled frames. Our proposed video feature extractor network, which consists of an image feature extraction network followed by a new transformer encoder, is trained end-to-end by including a classifier head that uses cosine similarity layer instead of the traditional linear layer to classify a corpus of labeled video examples. Unlike prior work in meta learning, we do not use episodic training to learn the image feature vector. Also, unlike prior work that averages frame-level feature vectors into a single video feature vector, we combine individual frame-level feature vectors by using a new Transformer encoder that explicitly captures the key, temporal properties in the sequence of sampled frames. End-to-end training of the video feature extractor ensures that the proposed Transformer encoder captures important temporal properties in the video, while the cosine similarity layer explicitly reduces the intra-class variance of videos that belong to the same class. Next, in the few-shot adaptation phase, we use the learned video feature extractor to train a new video classifier by using the few available examples from novel classes. Results on SSV2-100 and Kinetics-100 benchmarks show that our proposed few-shot video classifier outperforms the meta-learning-based methods and achieves the best state-of-the-art accuracy. We also show that our method can easily discern between actions and their inverse (for example, picking something up vs. putting something down), while prior art, which averages image feature vectors, is unable to do so.

Read more