Projects
Visual Embeddings and Re-identification: We demonstrate the importance of sampling informative data during the training of triplet loss embedding. In this work, we also introduce evaluations on a triplet-sampling-variant: "Batch Sample" into the re-identification literature.
Ratnesh Kumar et al. A strong and efficient baseline for vehicle re-identification using deep triplet embedding, JAISCR 2020.
Ratnesh Kumar, Edwin Weill, el al. Vehicle Re-identification: An Efficient Baseline Using Triplet Embedding, IEEE IJCNN 2019, ORAL. [pdf]
Co-author for CityFlow Benchmark at CVPR 2019 - ORAL. [pdf]
Co-author of PAMTRI: a pose-aware multi-task learning method for vehicle ReID using highly randomized synthetic data. ICCV 2019.
Chapter 10, Visual Embeddings in Deep Learning for Vision Systems. [LINK]
Exploiting Global Features for Robust Pose Tracking: In this work, we improve pose tracking using global color and motion features. Global motion feature computation is aided by extracting multiple "diverse" solutions from a model.
Multiple Object Tracking : In this work, we view multiple object tracking as a graph partitioning problem. Given any object detector, we build the graph of all detections and aim to partition it into trajectories. To quantify the similarity of any two detections, we consider local cues such as point tracks and speed, global cues such as appearance, as well as intermediate ones such as trajectory straightness. These different clues are dealt jointly to make the approach robust to detection mistakes (missing or extra detections). We thus define a Conditional Random Field and optimize it using an efficient combination of message passing and move-making algorithms. Our approach is fast on video batch sizes of hundreds of frames. Competitive and stable results on varied videos demonstrate the robustness and efficiency of our approach.
[MRF, Multi-labeling , Message Passing, Iterated Conditional Modes]
Video Representation : In this work we aim to join spatial and temporal aspect of a video into a single notion : Fiber. A Fiber is a set of trajectories spatially connected with a triangular mesh.
Pros :
1. A spatio-temporal neighborhood system to compute various relevant criterions.
2. Unlike approaches based on super-pixels, a fiber is built jointly by considering spatial and temporal aspects of a video.
3. In addition to providing association label for a pixel, we also provide a "long term" temporal reliability measure which assesses long term temporal color coherency.
Distance Correlation based Feature Descriptor : Traditional covariance measures only linear relationship between random variables. To overcome this, recently a novel measure is proposed by [Szekely & Rizzo 2009] using ideas stemming from the Brownian Motion. Distance correlation between two random variables is zero iff they are independent or the samples are identical.
Slawomir Bak, Ratnesh Kumar, Francois Bremond. Brownian descriptor: A Rich Meta-Feature for Appearance Matching, IEEE WACV 2014.[pdf]
More Information on Brownian (Distance) Correlation can be found on this webpage : Link
Insect Classification in Greenhouse : The purpose is to develop a classification system for the automatic detection of harmful insects in greenhouse plants. The classifier needs to decide among three classes : whitefly, greenfly or a background (i.e. false positive from the detection)
Tree Counting in Areal Images : The workflow consists of low pass filtering and anisotropic diffusion to smooth out small gradients and, subsequently detecting blobs using zero crossings of the second derivative in the diffused image. Its motivated from the work on olive-tree extraction in satellite images by [K. G. Karantzalos , D. P. Argialas 2004]
Diseased and Healthy Leaf Classification for Citrus Trees : As can be noticed in the following figure, the texture differs significantly for the diseased leaf. Hence the use of Haralick's features for classification.
C. Wetterich, R. Kumar, S. Sankaran, J. Belasque Jr., R. Ehsani and L. Marcassa. ”A comparative study on application of computer vision and fluorescence imaging spectroscopy for detection of Huanglongbing citrus disease in USA and Brazil” Journal of Spectroscopy, 2012. [pdf]